# 05 2026-07-18 Langchain Agents Intro Langgraph

course: Module 5 — AI Agents & Agentic Frameworks
module: Module-5-AI-Agents-Agentic-Frameworks
date: 2026-07-18
type: transcript
video_url: https://personal-learn.armco.dev/files/_Recordings/Module-5-AI-Agents-Agentic-Frameworks/05_2026-07-18_Langchain_Agents_Intro_Langgraph.mp4

---
[00:18:59] Hello, guys.
[00:19:05] Okay.
[00:19:09] Good evening, guys, good evening.
[00:19:15] Okay, so today we will continue on the evaluation.
[00:19:20] Last part is left.
[00:19:22] And uh…
[00:19:25] And then we will move on to
[00:19:28] Other LangChain components.
[00:19:56] Yeah, so last day.
[00:19:59] Uh, we were talking about ragas, and in the ragas part, last day we
[00:20:04] Discussed with custom LLM as a.
[00:20:08] As a judge. This is the last thing, Deepan, do you remember? This was the last thing, right? Or?
[00:20:14] Or was it?
[00:20:17] Or the last thing was this non-LLM evaluation.
[00:20:23] As far as I remember, custom LLM as a judge was done.
[00:20:35] Do y'all remember?
[00:20:39] Yep.
[00:20:41] Yeah, this was done, and then at the end I told you that I'll mix it with the non-LLM metric, and we will talk about the Raghas thing. So anyways.
[00:20:49] So, in the custom LLM judge, we created a pedantic architecture, so I had told you also that pydantic thing we will understand more today.
[00:20:58] Uh, and today.
[00:21:02] Yeah, today we'll see some, you know, some tool calling example over there. We will be more clear about Pydantic. But as of now, you just understand that.
[00:21:10] Your model will be forced to output.
[00:21:14] In this way only, they will be forced to output a JSON structure which has model config.
[00:21:19] which has correctness, relevance, fear, faithfulness, and reasoning.
[00:21:25] As the output, and then we give a prompt.
[00:21:28] And, uh…
[00:21:30] And during the model call,
[00:21:32] Uh, we mentioned we have a prompt of being a strict RAG evaluator.
[00:21:37] And we also mentioned we pass this schema as well, which we mentioned judge result. This thing we already discussed, judge result.
[00:21:44] As a schema as output schema.
[00:21:46] as a response format. And that's how we were getting
[00:21:50] the output in this format, and I had told you that many times people use.
[00:21:56] LLM not for LLM correctness.
[00:21:59] Relevancy or faithfulness.
[00:22:01] Okay, uh, usually people use LLM as a judge for
[00:22:06] LLM reasoning, ok, because reasoning helps us to take other decisions as well.
[00:22:11] Okay, so that was the reason, and then we joined it with the rest of the metrics, and this is how the overall.
[00:22:18] Thing looks like.
[00:22:20] Okay, reshell this collab link. Yes, sharing this. This is the same collab link that we were doing last day.
[00:22:41] Okay, here is the collab link that is shared.
[00:22:44] With all of you. Okay. Okay. So this is it. This we have now done.
[00:22:52] Now, I had told you that, uh…
[00:22:56] That we will be doing about, we will be talking about Ragha's LLM-based metric as well.
[00:23:01] Al-Ragas LM-based metrics. So, metric
[00:23:04] That uses LLM along with embeddings.
[00:23:08] for its own, uh, you know, calculations and all these things.
[00:23:12] Uh, so that we will be using today.
[00:23:15] Okay, so for that over here, see, this is where the code starts from.
[00:23:20] Okay, so Ragha's, you saw, we were using semantic similarity, then non-LLM similarity, string similarity.
[00:23:27] The Lebenstein distance thing and.
[00:23:31] All those things we saw. Now, there are these popular metrics that we use as evaluation that is known as faithfulness.
[00:23:39] We were already doing it. Without ragas, we were asking the LLM to come up with faithfulness, ok.
[00:23:44] But ragas also have faithfulness.
[00:23:46] Raghaza, faithfulness, ragas have answered relevancy.
[00:23:51] Answer correctness.
[00:23:52] Context precision, context recall.
[00:23:55] Okay, all these metrics are also there. These metrics are mostly LLM based. They actually, you know, wrote.
[00:24:01] Like, depends on LLM only.
[00:24:04] to answer it, because they have.
[00:24:06] intelligence involved in their calculation.
[00:24:09] Okay, so over here, I am doing the… I'm creating the Raghas Judge LLM, which
[00:24:16] has the LLM factory as a
[00:24:19] As a, you know, rapper.
[00:24:21] Uh, which takes in the LLM call and all those things, so all your open router client and all these things that you have created above will be taken, but
[00:24:27] One thing, if you give your
[00:24:30] Uh, you know, open router.
[00:24:32] If you give the open router access to Ragas, make sure you are giving the async open router, not the
[00:24:39] Not the synchronous open router, so let me just
[00:24:42] Uncomment this asyncout.
[00:24:44] And play this again.
[00:24:49] Okay, because I told you Raghas is an async.
[00:24:52] Ragha's calls are async in nature.
[00:24:54] So just this much we will have to do.
[00:24:56] Now you come down and then
[00:24:59] Now you, uh, play this, now this will take the async open router client.
[00:25:04] And provider is OpenAI, because your model is OpenAI, so if you have Claude models, then write Claude over here.
[00:25:09] Okay, so this, so internally, you know, internally, depending on this, you know, ragash changes its structure.
[00:25:19] Okay, so they have all these internal wrapper built out, this LLM factory, you know, take, take this as a parameter.
[00:25:26] That you need to mention, and judge model is the LLM model that sorry the.
[00:25:32] Judge model is this, actually.
[00:25:35] Great.
[00:25:40] Yeah, this is the LLM open router model.
[00:25:44] that you have already imported at the top.
[00:25:46] So judge model you pass.
[00:25:49] Yeah, judge model you pass. Now, let's look at faithfulness.
[00:25:53] Faithfulness, if you look,
[00:25:56] It is requiring that.
[00:25:58] LLM. Context position is requiring the LLM.
[00:26:01] Context recall requiring the LLM.
[00:26:05] Other than that, answer relevancy metric will require the LLM as well as
[00:26:10] The embeddings as well.
[00:26:12] Both LLM as well as embeddings.
[00:26:15] Answer correctness also requires both the LLM and the embedding.
[00:26:18] Okay, so let's go down and just look at the definition.
[00:26:23] Faithfulness checks whether the generated answer is supported by a retrieve context or not.
[00:26:29] So, your LLM
[00:26:31] Since faithfulness while checking faithfulness, you have the LLM access.
[00:26:35] Let's say your
[00:26:37] RetrieveContext is your… the OCR failure rate in European legacy branches is 15%.
[00:26:44] This is the context that you have retrieved.
[00:26:46] And the generated answer is the OCR failure rate is 15%, the hardware upgrade was completed last week.
[00:26:52] So, is it faithfulness, faithful or not?
[00:26:57] The first treatment is supported by a retrieve context. The second statement is not.
[00:27:01] So, you can say this is like 50% faithful. The answer is.
[00:27:07] The generated answer is 50% faithful because half of the answer, there is no retrieve context for it. It has…
[00:27:13] Road from its side.
[00:27:14] So this is where faithfulness will be lower in this case. A higher faithfulness code means that most of the answer is supported by retrieved information.
[00:27:21] Lower faithfulness score means the answer contains unsupported or hallucination information.
[00:27:27] Okay, so that is what faithfulness is. Second, after faithfulness.
[00:27:32] Uh, we have is context precision metric.
[00:27:35] Let's look at over here.
[00:27:37] Okay, after faithfulness, we have relevancy result. Let's look at relevancy result.
[00:27:41] Okay, answer irrelevancy. Answer relevancy, simple is checks whether the generated answer directly matches with the question or not.
[00:27:48] user question. So, question and answer similarity is checked, okay.
[00:27:51] So you can, you usually have an embedding model. Embedding model.
[00:27:55] can be easily used. So, embedding model, you pass your answer, you pass your question.
[00:28:00] It checks the distance, and from there, it tries to understand.
[00:28:04] Let's say what is the OCR failure rate? The generated answer is the OCR failure rate is 15%. Very to-the-point answer.
[00:28:10] This answer directly responds to the question, so answer relevancy score should be high.
[00:28:14] So now consider this answer. The scanner firmware is version 22.1. This statement may be correct, but it does not answer the question about OCR failure rate.
[00:28:23] Therefore, its relevancy score should be low in this kind of cases.
[00:28:29] Okay. So this is what answer relevancy is. Answer relevancy is more about if the answer is relevant to the user question or not.
[00:28:37] Okay. Third comes is answer correctness.
[00:28:41] Correctness checks whether the generated answer agrees with the ground truth answer or not. So there is a ground truth answer, that is our true answer, expected answer.
[00:28:49] Okay, that we have expected answer. That could be a human answer as well. This has nothing to do with context or anything. If you have ground truth answer.
[00:28:57] And if your answer, whether your facts in the ground truth generated answer are correct.
[00:29:01] Whether the overall meaning is similar to the
[00:29:03] Ground truth answer or not. So your reference answer or your expected answer, that is overall meaning is correct or not.
[00:29:10] With the LLM answer, or whatever answer you have got, maybe the top answer of your.
[00:29:15] of your embeddings as you answer. So, the judge model takes whether
[00:29:19] Facts are correct or incorrect or extra or missing or not. The embedding model checks whether the generator answered and the reference answer have a similar meaning or not.
[00:29:27] Okay, so let's say the OCR failure rate is European legacy branch is 15%.
[00:29:34] European legacy branches have 15% OCR error rate.
[00:29:39] Okay, so the wording is different, but both answers communicate the same fact.
[00:29:44] Therefore, answer correctness should be high in this kind of case. Okay, now consider the ODCR failure rate is 25%.
[00:29:51] The answer discusses the correct topic, though the value is incorrect.
[00:29:54] Therefore, fractual correctness code would be low.
[00:29:58] Okay, so the answer correctness code in this case will come down.
[00:30:02] Okay, so…
[00:30:05] This is answer correctness. The last 2 you have is context position and context recall.
[00:30:13] This is mostly with respect to
[00:30:18] uh, like, lesser with respect to answer.
[00:30:22] Uh, more with respect to the context that you have retrieved, so.
[00:30:26] This answer correctness, answer relevancy is very much LLM eval type of a thing. It's like deep eval.
[00:30:31] So, it is like how your answer is more relevant.
[00:30:35] with the user question. It has… it has lesser dependency, like, it has
[00:30:40] Like lesser direct check with the retrieve context. With the retrieve context, it is not checked. So, if you look at
[00:30:45] answer Levinci, did you pass retrieve context? No.
[00:30:48] You didn't you don't need that. You actually check with a ground truth answer.
[00:30:52] Okay, even same goes for answer correctness as well.
[00:30:55] Okay, but retrieveContext is used for faithfulness.
[00:30:59] And retrieveContext is used for
[00:31:01] Context position and
[00:31:04] Even recall as well, so these are more rack-specific.
[00:31:07] Evaluation technique. So you, many people had questionna that can we use this for LLM eval as well? Definitely.
[00:31:14] Answer relevancy, answer correctness then.
[00:31:16] semantic similarity which we had.
[00:31:18] Then non LLMs, string similarity, Rogue score, all of them.
[00:31:22] that can be used as a LLM eval technique, even a deep eval, these things are there.
[00:31:26] Okay, but other things like context position, all these things are very rack specific. This is like actual RAGs evaluation.
[00:31:33] Okay, context precision checks whether the retrieved document.
[00:31:36] Are relevant to the questions or not. So, let's say you have retrieved 3 documents.
[00:31:40] Are those 3 documents even relevant?
[00:31:42] Or are you creating some extra as well?
[00:31:45] Okay, there could be, let's say, 3 documents out of them, like, 2 are extra, like,
[00:31:50] Like we had that example of rabbit, rabbit being extra, dog being extra.
[00:31:54] Okay, it also checks whether the most useful documents appear near the top of the retrieve results or not.
[00:32:00] Okay, like, if the ranking of the retrieved results is good or not.
[00:32:05] So relevant documents appearing at the beginning are considered more valuable.
[00:32:08] Then, relevant documents appearing near the bottom.
[00:32:11] Okay, so ranking also matters.
[00:32:13] Ah, a relevant context about the OCI failure rate.
[00:32:16] irrelevant context about the scanner installation.
[00:32:19] And another relevant context about OCR errors.
[00:32:22] Okay, suppose these are three ranked responses you have.
[00:32:26] This retrieval result is reasonably good, because two of the three contexts are relevant.
[00:32:31] However, one irrelevant context appears before the second.
[00:32:35] Useful context.
[00:32:38] Okay, so this one appears in between. So that's where it fails.
[00:32:42] Therefore, the context position score would be good.
[00:32:45] But not perfect. Now suppose the results are relevant context, relevant context, and irrelevant context.
[00:32:51] This ordering would receive a higher context position because.
[00:32:55] of the useful information appearing first. So, when your order is important.
[00:32:59] And when you're in the order, if there is irrelevant.
[00:33:02] Things is coming or not is important or not, then.
[00:33:05] You require precision.
[00:33:07] precision. So precision is more about how precise your results are.
[00:33:13] Okay, how precise are your results at? If your results is giving something extra, then your precision will be lower.
[00:33:19] A high-context fiction score means retrieved answer mostly.
[00:33:22] useful documents and rank them well.
[00:33:25] So, that is what, you know,
[00:33:27] Context precisionness.
[00:33:29] Okay. Any question guys as of now, guys? All good? If it is good, can you give a thumbs up?
[00:33:35] How's out.
[00:33:38] Sir, where these scenarios can be used? I know the Rakavas is used for ragged
[00:33:46] to validation, how the quality in terms of… but how this course, how we can trust this judge model is doing character model
[00:33:56] How this model is?
[00:33:57] See, we are using the other, J model to validate the, in terms of fairness and in terms of relevancy context, whether it is giving accurate or not, correct?
[00:34:08] Mm-hmm.
[00:34:09] We are using the using the certain other LLM model, correct? But we do suppose we are doing correct, then we are evaluating with the ChatGPT
[00:34:20] So, how, it will give these
[00:34:21] Hmm.
[00:34:24] details, whether it is a,
[00:34:30] The relevancy context, our precision context, whether it is understanding correct or not, how we can validate in terms of percentages, because every model LLM will have a different standard because the crowd support the answers will be different, correct
[00:34:45] So, how it will be validated, how we know the LLM is superior than the base LLM which we are testing
[00:34:52] No, no, no, see, uh, the LLM superiority, itis checking. If you can move, go on mute, Prakash.
[00:35:01] Yeah.
[00:35:02] Yeah, uh, all these things…
[00:35:03] Evaluation is a very manual process, so once you do this, you manually check it also.
[00:35:08] So, you manually check it as in. It depends on what is the evaluation metric that you want to use. So, let's say.
[00:35:14] You saw that, for example, let us say the context position below.
[00:35:20] Score is coming very, very low, for example.
[00:35:24] So, now, why is it coming low? You will.
[00:35:27] dig a deep dive into it, why it is coming low, you will see that for some examples, so.
[00:35:32] Let me complete it, I'll answer this question. Let me complete it, because you have asked that.
[00:35:36] this question at a point where last point is left, that is recall.
[00:35:40] Uh, Aditya, do you have, I'll take the question once this is done, okay? Because these are questions.
[00:35:45] Yeah, which requires the last point. I thought any question regarding
[00:35:49] Till now, any confusion that only I wanted to ask.
[00:35:51] But yeah, so let's do recall as well. So recall, this is the most popular used
[00:35:57] metric, okay, in most of my RAG engines, I have used recall.
[00:36:01] Okay, I told you I have used it unknowingly and then later on Ragas came and
[00:36:07] Uh, and then I saw Ragas can also be used. So this is very
[00:36:10] This is like if your ground truth answer.
[00:36:13] Uh, let's say this is your ground truth answer. Context recall checks whether the retrieved documents contain all the information.
[00:36:20] required to produce the correct answer or not.
[00:36:22] Okay. And if you have extra answer, it doesn't matter.
[00:36:26] Okay, so it focuses on whether the important information was
[00:36:30] missed during retrieval or not.
[00:36:32] Let's say the ground truth answer is this.
[00:36:34] Okay, who's your failure peak on Tuesdays because of weekly bulk batch processing of.
[00:36:41] and written PDFs.
[00:36:43] This answer contained three important pieces. OCR failure peak on Tuesdays.
[00:36:48] Reason is weekly bulk bulky batch processing. The files are written on handwritten PDFs.
[00:36:54] So, are these three things…
[00:36:56] part of your context or not. Now, you might have in your context extra things as well. Doesn't matter.
[00:37:02] If that is extra, that precision checks that.
[00:37:05] But Recall doesn't check any extra. Recall checks.
[00:37:08] that if this is the ground truth answer.
[00:37:10] Okay, is my retrieval?
[00:37:12] Is having this information or factual information present or not.
[00:37:16] If the retrieve context mentions that failure peaks on Tuesday and the bulk processes is
[00:37:23] is the cause, but they do not mention the handwritten PDFs.
[00:37:26] Obviously, that means it has missed the
[00:37:29] miss the exact thing. So, just like.
[00:37:32] Let's say the expected in my example which I was giving you last week.
[00:37:37] If the expected response, expected URL,
[00:37:40] Uh, is…
[00:37:43] Present within the top 3.
[00:37:47] Uh, you know, responses.
[00:37:49] Then the recall will be high. This is known as context recall. The context recall will be high.
[00:37:55] Okay, that means. But…
[00:37:57] Along with that, there could be extra details as well. Like, obviously, out of top 3, one will be the relevant answer, other two might be extra.
[00:38:03] Doesn't matter. So in that kind of case, context recall is preferred.
[00:38:07] Okay, so that is known as context recall.
[00:38:09] That is what is known as context recall. So, this is the overall summary for all of the things that we have used.
[00:38:16] Okay, so if you go back now, and if you apply all of this, same way, just like you were doing semantic similarity and all.
[00:38:22] You'll, you will use the ragash await functionality.
[00:38:26] And it will call the LLM and get all the scores, and then at the end, you will have the overall.
[00:38:31] overall score as well. So, let the score come, and then we will talk about it.
[00:38:35] In the meantime, we can discuss from the previous output as well.
[00:38:38] So previously merged output. So, for example,
[00:38:42] Let's say, over here the answers are good, but let's say I am somehow
[00:38:49] Okay, just a moment.
[00:38:53] Yes, somehow, let's say I am not happy with, let's say this answer relevance is 0.86, it is not 0.86. Let's say the answer is coming as 0.5.
[00:39:02] Okay, for most of the answer, it is coming as 0.5.
[00:39:06] And I am not happy. Prakash, to answer your question. Point five.
[00:39:10] So now why it is not coming 0.5 is where.
[00:39:16] You will go deep dive into Prakash. So, you do not checks.
[00:39:22] You, you are over here.
[00:39:24] Checking the capability based on this number. So you, your domain expert,
[00:39:29] Or in the most of the cases, domain expert, if domain expert doesn't give you this kind of a golden dataset, you only create it yourself.
[00:39:36] So now you creating yourself might not be that good. Why am I telling you.
[00:39:40] Because when I was building these RAG engines at Oracle, I didn't understand most of the
[00:39:46] software bugs, terminologies only, because it is something, like, they have been using for a long time.
[00:39:52] They had some, you know, complex terms known as DST, VOS, space management, transaction.
[00:39:58] Okay, these are very complicated terms they had. They usually understood very well, okay.
[00:40:03] So domain expert. So what did I do? I created some queries just by randomly taking some random.
[00:40:09] queries. So what did I do? I ran a full
[00:40:13] Python script.
[00:40:15] Along the.
[00:40:18] these PDFs and I, I took out some 10 words, 5 words, 6 words queries from there.
[00:40:23] Okay, uh, and those queries…
[00:40:25] I used again a LLM only and changed it, and I created a golden dataset like this of 50 to 100 queries.
[00:40:32] And wherever I picked it up from, that became my expected answer, expected URL from where I picked it up from.
[00:40:39] Because I have the URL stored of each of the PDFs.
[00:40:42] So I made that expected URL. It was a very easy to create a golden dataset for me.
[00:40:47] Okay, but later on I found that some of the query that I picked up, because since I am writing a Python script to pick up random queries.
[00:40:53] Some of the queries are too random also in nature.
[00:40:57] So due to that, my answer relevancy was down because some of the queries might be
[00:41:01] This error is used to solve.
[00:41:05] A space-related problem in Oracle, very general concept.
[00:41:10] Okay, let's say I have picked up that also as a query and I have kept that as my golden dataset.
[00:41:15] So this kind of, you know.
[00:41:18] When you are picking up this kind of sentences, your sentence might be too generic also in nature. So I realized that my answer was not coming good further. Then I went to domain expert, domain expert gave me some better queries.
[00:41:30] And that's how we came up with the 50 queries.
[00:41:32] Now, let's see your answer is not coming good.
[00:41:35] If your answer relevancy are not coming out. So, why video answer relevancy will be bad? Why do you think? Prakash?
[00:41:41] Why do you think your answer relevancy can be bad? So, if you look at answer relevancy, the definition.
[00:41:47] This is the definition of answer relevancy that whether
[00:41:49] Your answer directly answered to your user question or not, right?
[00:41:55] So, Prakash, now tell me, while looking at this, if your answer relevancy, average score of your answer relevancy is very bad, let's say 0.5.
[00:42:02] Now, how do you, what other reason that can be it is bad for?
[00:42:11] Hello.
[00:42:13] Yeah, it may be,
[00:42:16] Relevant checks are as not appropriate how we thought.
[00:42:19] Yes. So, you will go one level deep now.
[00:42:22] You will start printing the relevant chunk as well.
[00:42:25] Okay, so you will see if your answer relevancy is bad, then the checks that you have done with chunks is also not good.
[00:42:32] Okay, maybe your recall will also be lower. Your precision might also be very bad. You will see a very.
[00:42:38] Chain of metric being failing.
[00:42:40] Okay, with this. So you might go and ultimately you will come to the last point, that is, your chunks retrieval is only not good.
[00:42:48] So, now…
[00:42:50] If this kind of problem.
[00:42:52] happens. For many of the answers, many of the answers.
[00:42:57] Then you will go again.
[00:42:59] And maybe, you know, you will take some strategic decision.
[00:43:02] Let's say you might go for a different chunking strategy. You might not.
[00:43:07] Use the current separator, you know, by default separator, which is there in Raga, uh, which is there in.
[00:43:13] recursive character text splitter, you might not use that.
[00:43:15] So you will go for a different chunking strategy. You will have to think around.
[00:43:20] your use case that time, that, am I happy with the semantic results or not? Then.
[00:43:25] Uh, then you might have a hybrid re-ranking strategy also. So all the things we will start to question.
[00:43:32] Okay, that why is it not coming?
[00:43:34] Okay, in our case, what did we do? We were not happy with the top two result, uh, top 5 results first.
[00:43:40] Because top 5 results was giving very irrelevant answer as well.
[00:43:44] Because our, you know,
[00:43:46] Uh, there are chances in that.
[00:43:48] 400, 500 URLs that we were processing for particular department.
[00:43:53] There are chances there are too many.
[00:43:56] Each of the URLs are disconnected to the other URLs a lot, so
[00:44:00] There are chances of getting irrelevant.
[00:44:02] URL being returned. Okay.
[00:44:04] So the point is now.
[00:44:07] If from this irrelevant URL.
[00:44:09] you will have a relevant answer, and you… your token cost will also go high, and your answer will be relevant as well.
[00:44:17] So what…
[00:44:19] Will you do? What can you do over here?
[00:44:22] is you can trim down to top 3.
[00:44:25] Okay, uh, you can trim down to top 3 first of all. You can also do one more thing.
[00:44:30] When you are responding,
[00:44:32] Uh, in my case, what did I do is when I am responding from the
[00:44:37] You know, when I'm asking the question and I'm getting the answer.
[00:44:40] I was getting top 10, top 20 answers because
[00:44:44] See, guys, one URL is divided amongst multiple chunks.
[00:44:48] Okay, so if you do top 20, it will give you top 20 chunks, not top 20 different URL.
[00:44:55] So, from top 20 different URL, if then another logic is written on top of that.
[00:45:00] That if majority of this URL is present in most of the chunk, majority is chosen.
[00:45:05] Okay, majority is chosen over.
[00:45:08] Uh, now if the majority is all, like, let's say out of this 20, all are 4, 4, 4, 4, 4, 4 chunks are there so there is no majority.
[00:45:15] In that case, we prefer the score.
[00:45:18] So all these things are, it's your own IP or it's your own logic that you write on top of the retrieval.
[00:45:25] Okay, based on this evaluation, so
[00:45:29] After this evaluation, there are lots of steps that you can think of.
[00:45:32] Okay, in my case, this suited better.
[00:45:34] Where you go get more answer retrieved from there.
[00:45:38] You only take the majority ones as your…
[00:45:41] You know, uh, top retrieval chunks, top retrieval chunks, if the majority ones are
[00:45:45] Uh, if you cannot find the majority one, just take the highest ranked one. That is one of the preferences.
[00:45:50] And if, ah.
[00:45:53] That is 1 of the, that is 1 of the way now if you are not happy with the retrieved responses from the semantic search.
[00:46:00] have a BM25 as well, have a full-text search as well.
[00:46:03] Okay, or Elasticsearch as well. From there, also, you can get it.
[00:46:06] Okay, that is one of the ways. Now,
[00:46:08] These are the things that, you know, we do. Like, this is very open-ended that you apply. That is, like, when you are doing that problem.
[00:46:16] These are the things that you should explore.
[00:46:19] Now, you're not there could be other chances as well. There could be, let's say, your embedding model is not good.
[00:46:23] And then you're not happy with sentence transformer at all. There are people who are not happy with the sentence transformer.
[00:46:30] So you can that time, you know, sentence transformer is the top, okay, the best that you can achieve from the open source models.
[00:46:36] Then what you will do, you will go for OpenAI embeddings, you will pay for the OpenAI embeddings, go for a bigger size embeddings.
[00:46:42] OpenAI embeddings give you 300 size, 300 plus size embeddings.
[00:46:46] You will go for those. Okay, you will go for Gemini embeddings, you will try with those embeddings as well.
[00:46:52] So these are the things that we all do.
[00:46:55] Cottage Pragash.
[00:47:01] Yes, yes, sir.
[00:47:02] Yeah, but that's what I'm telling. See, these things are very non-standard method. Like, non-standardizing, there is no standard, what you will do.
[00:47:09] During your case, what will be there?
[00:47:12] It's very difficult to say, but most of the things are revolving around these things only.
[00:47:16] You are most of the, you know, answers will be around this only either.
[00:47:20] You will change your some of your retrieval engine.
[00:47:23] Okay, maybe I told you, my retrieval engine was not plain. It was not like I'm getting just stopped.
[00:47:28] three answers that I am showing. I am getting top 3 URL.
[00:47:31] Not top 3 chunks.
[00:47:33] from if you get top 20 chunks or 30 chunks from there, you will get top three URLs.
[00:47:39] Because chunks might be coming from the same URL multiple times.
[00:47:43] Okay, different, different chunks might be coming. If you ask too many too relevant questions that is present in multiple chunks.
[00:47:51] It can come. So…
[00:47:53] It was not a straightforward thing, so.
[00:47:54] Now, this kind of logic that you apply on spot is something that will be very specific to your use cases.
[00:48:00] Okay, so…
[00:48:02] That is the point. Yes, Aditya.
[00:48:05] You had one question.
[00:48:06] Yeah, can you go to the table, uh, in the bottom, where you're comparing all the metrics?
[00:48:11] Hahaha.
[00:48:14] Yes.
[00:48:15] No, not this. In the explanation, you were… yeah.
[00:48:19] Oh, here?
[00:48:20] Yeah, yeah, this is fine. Uh, so context precision is when relevant documents are there, retrieved, and context recall is when they have all the information, right? Are those not related to each other?
[00:48:31] They are very much related to each other, but they are confusing as well. Do you, do you
[00:48:35] Do you, in general, know what is recall and pre?
[00:48:38] Yeah.
[00:48:39] Uh, with respect to machine learning, you know, right.
[00:48:41] Yes, yes. Yeah, yeah.
[00:48:42] ML? Yeah, so if you look at this, this is that only. If you just take there is a very like, you know, while explaining there is a slight difference.
[00:48:50] Okay, uh, over there, it was purely mathematics, and that's how it is. Over here.
[00:48:55] If you look at it, the recall.
[00:48:58] Can have extract information.
[00:49:01] Okay?
[00:49:02] The recall will still be high, it doesn't matter.
[00:49:05] It checks from the point of view that whether the retrieve context is there or not, let's say.
[00:49:10] Let's say my goal is…
[00:49:12] I am choosing a football team.
[00:49:15] Hmm.
[00:49:16] From here, from this entire class.
[00:49:18] Now, my goal is that whether from this batch,
[00:49:23] Aditya.
[00:49:26] Prakash and Gunjan is there or not.
[00:49:29] Yes.
[00:49:30] Okay, this is my preference. Now, with that extra people are there, that is fine.
[00:49:32] Yeah.
[00:49:33] So, it is like that, recall is like that. If you have extra information is filed.
[00:49:37] Precision penalizes extra information.
[00:49:40] Okay.
[00:49:41] Okay, uh, by penalizing means Precision will prefer.
[00:49:45] that you shouldn't have irrelevant context coming up. So, like, rabbit and
[00:49:50] See, CAD was coming, along with that rabbit and dog was also coming, right? So precision would be low if you have that.
[00:49:56] But…
[00:49:59] But instead, if you do top 1 over there,
[00:50:02] your contention will be high in that case.
[00:50:05] Okay, in context of RAG especially,
[00:50:08] Can you think of a use case where precision is high but recall is low, or vice versa?
[00:50:14] Hmm, see, I can tell you the use cases where each
[00:50:18] will be required. See, when context, when
[00:50:22] Your search response.
[00:50:24] is not at all valuable.
[00:50:27] Hmm.
[00:50:28] Okay. Search response as in your context, that is the ranking of the context is not at all valuable.
[00:50:33] It should be there. That answer should be there.
[00:50:37] Okay. In that kind of case, you will go for recall. You will not think about precision at all because you will die.
[00:50:44] You know, fixing your position.
[00:50:46] Diaries in your project will die.
[00:50:48] Hmm.
[00:50:49] fixing or precision problem. Okay.
[00:50:50] Okay.
[00:50:51] Where you just have the answer should be coming.
[00:50:54] Okay, but you might have extra answer, two, three answers extra, no problem.
[00:50:59] In that kind of case, recall is
[00:51:03] Uh, preferred. Okay, like in my case.
[00:51:06] Uh, I always use Recall because ultimately I am showing top 3 to top 5 answers. So
[00:51:11] Usually people check those search response from those top 5 responses.
[00:51:16] And they choose their relevant document.
[00:51:18] Now, look at the problem that I'm solving.
[00:51:20] Previously,
[00:51:22] What are my problem? The problem I'm solving is software bugs. Software bugs, people are stuck in software bugs for weeks before solving it. Why? Because they're not able to go to the right document only.
[00:51:33] There are millions documents in.
[00:51:36] Oracle. Okay.
[00:51:37] Now, the problem is they have a search mechanism. Oracle has one of their own search mechanism.
[00:51:44] Which checks around this million documents before coming to the right document.
[00:51:47] Okay, now I am making a department-based chatbot. So obviously my search space will reduce. Okay.
[00:51:56] And it'll be more focused.
[00:51:58] Yeah.
[00:51:59] Now, over there, if I can just show the relevant answer, along with that, if I show 5-4 extra, it's not a problem.
[00:52:05] Because, look at the problem. Before they were going through millions of documents to find the right answer.
[00:52:10] Now they are going through five documents to find the right answer.
[00:52:14] Okay, yeah.
[00:52:15] Got it. So, your
[00:52:16] Recall is important, ok, just a moment, just a moment, recall.
[00:52:22] Somebody's not.
[00:53:02] Okay, now coming to…
[00:53:08] Uh, this is with recall to, uh, this is with respect to recall. Understood, Aditya, recall.
[00:53:12] A concept?
[00:53:13] Yeah, yeah, yeah.
[00:53:14] Now the point is…
[00:53:16] Consist, uh, context position is more has to do with sellability.
[00:53:22] When your relevancy of the rank is very important.
[00:53:25] Forget the LLM part, ok, LLM, so I'm not considering only as of now.
[00:53:29] Let's say you have a ranking.
[00:53:33] Based on that ranking, people will move ahead, or they will stop using our product.
[00:53:37] Okay.
[00:53:39] In that kind of case, where
[00:53:42] you know, user experience is important. Now, this is something that is coming at the top of my mind, as one of the use case.
[00:53:48] Where your
[00:53:50] Precision has to be good, because if you give irrelevant response.
[00:53:54] Irrelevant ranks of response also, then people will stop trusting you.
[00:53:59] Okay, for example, for example, I'm giving you a very
[00:54:04] Easy example that comes to my mind is, let's say Zomato. Tomorrow I have an issue with the food.
[00:54:09] And tomorrow I go to Romatos chatbot and I ask that.
[00:54:13] uh, some issue with my food and
[00:54:15] Zomato, from its set of Q&A, it just gives me some relevant question.
[00:54:20] Now, relevant answers. So, is the food has a spillage issue.
[00:54:23] Burnt, and all those things. And let's say my
[00:54:26] Inquiry is not there only.
[00:54:28] The kind of issue. It's say I have some problem with the delivery.
[00:54:31] Okay, and that issue is not there only. So I'll feel frustrated.
[00:54:34] Okay, naturally, so…
[00:54:37] When you have the sellability concept, where the users
[00:54:41] where your ranking is very important. In that kind of case, precision is very important.
[00:54:45] Yeah, makes sense.
[00:54:48] Yes, Gunjan.
[00:54:49] Yeah, to support your point, right, what to ask, like, okay, for the use case where that
[00:54:55] We need a high recall, low precision, right?
[00:54:58] So just give some example.
[00:55:01] Uh, we are facing daily vision on the enterprise level, that
[00:55:04] Example audits are there, right? So when something audits are happening at the enterprise level, right?
[00:55:09] We want to collect multiple evidences, right? We cannot…
[00:55:13] Filtered out at the early stage, we want to collect each and every artifact what is available, which is known, right.
[00:55:19] So example, quarterly we are doing audit of, like, SOC audits are happening, the regulatory audits are happening, right?
[00:55:26] And in some cases where that, um,
[00:55:28] lawful activities are happening, right? If you have heard about that, okay, a lot of, like, courts are putting some fine.
[00:55:35] Or penalties on the organization, right, doing some activities, right?
[00:55:38] So, in that case is while doing audits, right, so especially the
[00:55:42] Uh, the, the security or the auditors.
[00:55:45] would like to collect each and every evidences what is linkage. But there is a correct or not, that is secondary strain that will be filtered out, the human is going to look for that, right?
[00:55:54] But in that case, focus is to collect more and more evidences which is linked to the particular.
[00:56:00] But I would say potentially linked to that.
[00:56:03] lead to that particular event.
[00:56:06] Right? It might be ignored later stage, but we need to collect this data. So this is a regular, I think we all are doing like enterprise like.
[00:56:07] Mm-hmm.
[00:56:12] Audits are happening quarterly, financial audits, regulatory audits.
[00:56:16] privacy audits, multiple companies are doing some penalties or suing to each other companies, right?
[00:56:21] So, that case is recall is…
[00:56:23] High precision might be low will suffice.
[00:56:25] Yeah, yeah, because you might have…
[00:56:26] Right, and which is a very practical case and I think, yeah.
[00:56:29] You might have irrelevant things as well in coming in the audits. Yeah, yeah.
[00:56:38] Hmm, hmm.
[00:56:39] Yes, yes, we are collecting evidence is what is required just to prove our points, then as a human they can look for the others actually needed, not needed, they can filter it out.
[00:56:43] Huh.
[00:56:44] That is the secondary stage. But that is something required. So that is where the high recall is required, low precision will suffer, ah.
[00:56:48] Well, okay, in that case.
[00:56:50] Yeah.
[00:56:52] Got it.
[00:56:53] So, yeah, so basically like from my experiences, just to add on top of that.
[00:56:59] Uh…
[00:57:02] From my experiences, I have seen, like
[00:57:04] Internal products are mostly high recall.
[00:57:06] High recall product.
[00:57:08] Like something shouldn't get…
[00:57:12] shouldn't get waved off.
[00:57:14] Okay, something that shouldn't get weaved out. Everything should be caught.
[00:57:18] Now then we take human, human will take the decision. That's the thing. So over here also, I'm doing the same thing. Top 5.
[00:57:24] The sponge should be there. Okay, now 3 might be relevant.
[00:57:28] Okay, but, you know, see, if I
[00:57:30] Due to this, due to fix, due to, you know, due to precision, if I miss out.
[00:57:35] On one of the important documents, I'll wait another 3 weeks to solve this software bug, guys.
[00:57:39] You all might be thinking that software bug is not a big issue, maybe because you all might be working on the other side of the thing.
[00:57:46] But I'm telling you, like, people who have built big products, like Atlassian, Oracle, who have built.
[00:57:52] Big products. Software bug is itself is a department.
[00:57:57] In many companies, it is known as sustaining engineering.
[00:58:00] Okay, over there, apparently Oracle used to lose around.
[00:58:05] uh thousand dollars okay uh as far as I remember it was 1,000, now it is around $1,200.
[00:58:11] per bug that used to get filed by a customer. Okay, this is some mathematical metric that they have found out.
[00:58:18] Okay, that every bug, like, out of, like, 4,000 bug that always used to lie on there.
[00:58:23] bug repository. Okay, they used to lose around $1,000 per bug, okay.
[00:58:27] So that point of time, now it could be 1200, 3 or 1300.
[00:58:31] Okay, so because this bug
[00:58:34] Used to take a lot of time to solve, one year, 2 years.
[00:58:37] Okay, and these people require the resolution faster, and they have millions of documents, apparently.
[00:58:42] It's just like, you know, they have their own stack overflow only.
[00:58:45] Okay, so anyways, so in my case, context recall has been the highest, most because I have mostly built
[00:58:52] uh, you know, products that is, you know.
[00:58:54] chatbot that is internal to that company, proprietary to that company, over there, RAG.
[00:58:59] And I'm being, at least the answer should be there, okay, in the top three response. Order ranking was not important.
[00:59:06] Ranking is more important when it is more retail phishing.
[00:59:08] Okay, where other people, external people are using it.
[00:59:11] Okay.
[00:59:13] Yeah, Anitman, just one follow-up. How do you calculate precision and recall in production system?
[00:59:18] Uh, in production system.
[00:59:21] See, what we do is you have the model observability, right?
[00:59:25] Hmm.
[00:59:26] So in model observability, ah.
[00:59:29] What do we do is sometime these
[00:59:33] People.
[00:59:35] We have the particular inquiry questions and their responded answer. This is what model observability does, right? Every question and their
[00:59:42] answers what is the expected answer.
[00:59:44] Uh, what's the answer that was given by the LLM? Let us say.
[00:59:47] Our system ran in the production for 6 months.
[00:59:50] Okay, and we have, we have collected in our DB.
[00:59:55] All the user inquiries.
[00:59:57] and their responses. That is there.
[01:00:00] Now, you can take, like, 20 from there to 200 from there, let's say, not 20, 200 from there.
[01:00:06] Hmm.
[01:00:07] and do the same process.
[01:00:09] of manually checking it. What manually checking it, as in the manual part over here is
[01:00:15] Build the same, uh, pipeline, run it through the ragas, those 200 enquiries.
[01:00:21] Hmm.
[01:00:22] in the production that you have collected from model Observability DB.
[01:00:26] You collect it, run it through your ragas. And if your ragas are too off,
[01:00:30] Okay, two off. Then you do an investigation.
[01:00:32] What is Aditya?
[01:00:33] Got it, yeah.
[01:00:34] So, let's say you are using the product. Let's say Zomato, I am using it, okay? Like, that first 200 users are using it, 2,000 users are using it.
[01:00:41] For after 6 months, I take this 2000 users ka input.
[01:00:46] And what did my LLM give? I collect that.
[01:00:48] Okay. And and all the retrieve context, everything I collect. I run it through Ragas.
[01:00:53] Same thing, same process.
[01:00:55] Okay, now the manual part, where is the manual part? If the answer is not good, if the relevancy is not good, then I will go one level deep.
[01:01:01] Then, bro, your golden dataset was working good, but in our production, many, many times it is not working.
[01:01:07] Hmm.
[01:01:08] That time, we'll have to check. That is.
[01:01:10] Okay, this will do all the analysis, right? You'll look at them particularly, you'll look at…
[01:01:14] Okay.
[01:01:15] Yeah, yeah, yes, yeah, yeah, that's where human thinking, human intelligence, and all those things, your
[01:01:17] knowledge, how what you have learned till now. Maybe by that time, better metric came.
[01:01:22] Okay, it could be possible AI is changing left, right, and center within few days, it is completely shifting.
[01:01:28] Okay, you might come up. Okay, our actually we applied, I told you, when I started, Raghas was not there.
[01:01:33] I applied context recall from our own knowledge, ok.
[01:01:36] by taking decision with… then, after 6-7 months,
[01:01:41] Ragas came in and they told that we need more metric and they applied 3, 4 more metrics as well using Ragas, so
[01:01:47] That, that happened due to the production thing only, like in the production, many.
[01:01:51] Many people in many teams, certain queries are not being answered.
[01:01:56] Okay, so so this.
[01:01:58] You know, I have been continuously telling about that I built a semantic search, and then I built a full-text search. This full-text search was built after few days in production.
[01:02:06] Okay.
[01:02:07] First, it was only semantic search.
[01:02:09] Then from the response we got that, no, enough, it is not enough.
[01:02:13] The answers are not coming. Then we build a full-text search.
[01:02:16] All these things, so there is a lot of manual intervention or you can say human in the loop.
[01:02:20] That happens in between when during evaluation and during this production stage.
[01:02:29] Okay.
[01:02:30] Okay, so anyways, guys, so that's… that is about Ragha's evaluation. Now, a lot of rags we have done in the last.
[01:02:37] I would say 3.5 classes.
[01:02:41] we have done. So, now, after this, probably we'll face RAG in agentic rag.
[01:02:46] It is just, like, you know, RAG becomes one component over there.
[01:02:51] Okay, it is not very complicated. It's very simple.
[01:02:53] Okay, now we'll go back to our LLM, basic LLM again, you know, talking about tool calling, talking about learning about.
[01:03:00] Uh, you know, agents and all the language chain agents as well, then we'll even eventually move on to LangGraph as well.
[01:03:06] So that is the point.
[01:03:09] Okay, so before going into agents, before that we have
[01:03:16] Another thing.
[01:03:19] Which is known as external tool calling. How many of you don't know about tool calling at all?
[01:03:26] Can you just raise your hand?
[01:03:29] I don't…
[01:03:32] Okay. Just give a hand raise, not thumbs up, hand raise if you don't know it, I just want to see.
[01:03:39] Like a good amount of people not knowing it.
[01:03:42] Keep it raised, ok?
[01:03:45] External tool calling.
[01:03:49] K4.
[01:03:52] Whatever the rest, all people know.
[01:03:57] Even, I don't know, I could not see the hand raise symbol, sorry.
[01:04:01] Okay, okay, okay, so many people could not be seeing it, so 8.
[01:04:05] Okay, good amount of 10, 13.
[01:04:08] God, guys, 14.
[01:04:11] Okay, so yeah, I got it. So, many people don't know it. So yeah, so first we will understand tool calling, we'll understand Pydantic.
[01:04:18] And then we'll go into LangChain agents, okay?
[01:04:21] So yeah, fine guys, I understood.
[01:04:24] Okay, uh…
[01:04:29] Okay. Okay, so this is something.
[01:04:33] Known as tool calling that you all should know, this is.
[01:04:39] In agnostic to any framework.
[01:04:41] Okay, we are just using plain Python.
[01:04:45] and OpenAI SDK in order to do tool calling. Okay, there is no LangChain, nothing we are using.
[01:04:51] Okay, uh…
[01:04:52] I will share this…
[01:05:06] See, if you are able to access it.
[01:05:08] Uh, can you share the previous one also, Ragha's sheet? I joined late.
[01:05:11] Yeah, yeah, yeah, yeah.
[01:05:13] Not sure if you have already shared.
[01:05:17] So, anyone, so whatever link you are sharing, you are giving that one to… that is being updated in the portal also, right?
[01:05:23] That will be shared in our different way to portal, it will be shared… the notebook will be shared with them.
[01:05:31] Okay, that I'm not getting from this.
[01:05:32] Okay. Yeah, it will be shared. It will be shared, but it will not be in a link format. It will be in a notebook format. Like, I'll just download it and
[01:05:38] I'll share it.
[01:05:41] It's accessible, uh…
[01:05:51] Yeah.
[01:05:54] Okay. Everybody able to open it?
[01:05:59] Okay, so guys, before me reading it.
[01:06:02] I want you all to once go through it.
[01:06:06] Uh, once… have a look, because that will, you know, ease the, you know,
[01:06:13] me, you know, scrolling up down like this very easily. Don't play anything as of now.
[01:06:17] Uh, don't have to play anything.
[01:06:19] Uh, because, uh, there is a limitation on these APIs as well, guys, because I have this weather API, I think weather API.
[01:06:25] You can make two requests per day or per minute, or something like that is there.
[01:06:30] So if you make it guys your by the time I start teaching, you will, your request will be over.
[01:06:35] Okay, so don't play anything. As of now, let me give you an idea what we are doing. So, first of all,
[01:06:41] Guys, you all have to go to weather, open weather map.
[01:06:44] Okay, so like that, there are lots of these APIs, okay? There is open weather map we are using. You can use OpenMatio.
[01:06:51] then there is something known as, ah.
[01:06:54] And I'll…
[01:06:56] Weather app as well, uh, many, many APIs are there like this.
[01:07:00] to get weather-related data and hourly forecast and all these things. So if you go to open weather, you can sign up.
[01:07:07] Uh, using your Gmail, or first you all do this, okay? You all do this.
[01:07:11] And you go to get API key. While signing up, you'll have to go to your Gmail and you'll have to do the verification, everything complete, ok.
[01:07:18] If you go to get API key, if you search that, you will come to this screen.
[01:07:21] And you just click API keys.
[01:07:24] Over here, create an API key.
[01:07:27] Give a name and create an API key generate if you press generate.
[01:07:30] Okay, by default, this will be there.
[01:07:32] Okay, I have never used the default one, I've always created, give a name, created, another entry will be created like this.
[01:07:38] So, copy this and keep it saved somewhere here.
[01:07:44] Okay, keep it saved somewhere here.
[01:07:47] Uh, as a weather API, like, like I have saved it as a secrets.
[01:07:52] Keep it saved because you will be requiring it. So, this is the weather API now.
[01:07:56] The point over here that we are trying to do is how do you
[01:08:00] Do simple tool calling. The tool calling means you might, your company might have a different API.
[01:08:05] Okay, your company might ask you to go through a, let's say, a news API.
[01:08:10] Your company might ask you to go some internal APIs as well.
[01:08:15] And using that API, you process the data. So that is why you need to understand the concept of tool calling.
[01:08:21] Okay, tool calling is connecting to external tools, APIs like that. Okay, and we are doing over here is weather tool over here.
[01:08:28] Later on, we will do like more tools as well, like now weather tool will add its search tool as well.
[01:08:33] So I'll mix it together and you'll do it.
[01:08:36] But this will be replaced by your enterprise-level tools that you will have.
[01:08:40] Okay, so that is what first you get this weather API
[01:08:45] Keys, guys. Check it out.
[01:08:47] And let me know if it is done.
[01:08:57] Hands are still raised, it is because of the last hand raise.
[01:09:03] Can you share the website link from where we need to get them
[01:09:05] Oh yeah, yeah, yeah.
[01:09:08] Okay, uh, I think, guys, can you put down your…
[01:09:11] Hand raise because it's confusing.
[01:09:14] Actually…
[01:09:25] Yeah, Kishore, can you put it down or do you have any question?
[01:09:33] Okay. Okay, guys, I have shared you the link.
[01:09:37] Can you all sign up and get the API key?
[01:10:16] Once you are done, guys, start writing done, because this API key, somebody, some few people feel find it difficult in getting it.
[01:10:22] Okay, don't start using it immediately.
[01:10:28] Because you have some limitation in these APIs, wait, I will tell you the limitation as well.
[01:11:13] Okay, so hourly forecast, all these things are available, calls per minute is 60, 60 calls can be.
[01:11:20] can be used. Okay.
[01:11:22] Hopefully, it is true to whatever it is mentioned, because I have seen.
[01:11:27] Most of the time, like, many people find it difficult.
[01:11:30] A few people will, you know, might find it difficult that
[01:11:35] your API call might get over.
[01:11:37] Okay.
[01:11:44] Yeah, 60 calls per minute.
[01:11:46] Okay, hopefully.
[01:11:48] So, Anir, when we save it in that, uh, Secrets Voila section, is it?
[01:11:50] Yeah, yeah, yeah, yeah, yeah.
[01:11:51] And any particular name to be given, or we can keep it whatever we want.
[01:11:53] I don't know, any name, any name, any name.
[01:11:56] Any, because anyways, we are…
[01:11:58] Over here, we are asking the user only to enter it. So you can just copy and paste it.
[01:12:02] Okay.
[01:12:03] If you wanted to get extracted from here,
[01:12:06] Then you can follow that naming convention.
[01:12:09] Okay, that also you can do, but then you'll have to change the lines a bit.
[01:12:13] Okay, wherever it is asking, you'll have to change it.
[01:12:16] From getpass.getto, userData.get, and you have to import user data as well.
[01:12:22] Of, but better keep it for now, you keep it like that only.
[01:12:26] Okay, as it… yeah.
[01:12:27] Okay.
[01:12:31] Yeah, it's done.
[01:12:34] Okay.
[01:12:37] Guys, if it is done, write down in the chat it's done so that I am.
[01:12:41] find that you all are done.
[01:12:49] I guess chat is disabled.
[01:12:51] That's why I'm not getting any answers.
[01:12:54] I'm assuming, because for me, it is showing us disabled, so…
[01:12:57] is what I feel.
[01:13:00] It's a disability.
[01:13:01] I gotcha.
[01:13:05] Okay.
[01:13:08] Personally, you can send me.
[01:13:10] Because as others have sent…
[01:13:13] Yes, we can send that.
[01:13:15] Yeah, yeah, done, okay, okay.
[01:13:18] Done, done. See, if at a good amount of people are done, then
[01:13:21] We'll start you can you, you can start going through the code once.
[01:13:27] Okay, you can start going through the code once yourself.
[01:13:30] What we are doing, it should be simple, guys. It's a very simple tool call only.
[01:13:34] Like it's a precursor to what we'll do next.
[01:13:59] I got 3.
[01:14:01] Seven people saying yes as of now. Done.
[01:14:06] Basically, seven people thing done.
[01:14:09] And rest are the people who have done it, guys, you can go ahead and read the entire thing that is happening.
[01:14:22] Try to understand what is Pydantic and all these things you will.
[01:14:25] you know, see that there is something known as pydantic we are following.
[01:14:29] See, like, this is the pyrantic structure that we are following. This is a…
[01:14:33] The base model will discuss on all these things.
[01:14:39] Okay, great. Eight people are done.
[01:14:56] Would 9 people done.
[01:15:31] Hello
[01:15:33] Okay, no, I don't have the open router API
[01:15:36] I'm not sure, like, whether maybe in the previous session, I must have created, but I haven't saved it yet
[01:15:37] Yes, yes. Oh, okay, okay.
[01:15:43] Go to Open Router.
[01:15:45] Okay.
[01:15:46] Uh, again, same, same, same option, get API key.
[01:15:50] Okay, if you have signed up, if you have signed up.
[01:15:53] Okay.
[01:15:54] new key.
[01:15:55] Give the name, okay.
[01:15:58] Okay, okay.
[01:15:59] Give the name and…
[01:16:00] Hello? How's up?
[01:16:01] Just create key, copy because it will not be…
[01:16:04] My dad
[01:16:05] It will not be shown like this.
[01:16:06] I'll be done equally.
[01:16:08] Somebody…
[01:16:10] Owen, can you please go unmute?
[01:16:11] Nobody, nobody.
[01:16:13] Awan, can you go on?
[01:16:15] I mean…
[01:16:16] Monday, and Tuesday, yeah
[01:16:18] Yeah, so Deepam, uh, like this, the key is visible in case of open router, it will not be visible, so please copy it and save it in your secrets.
[01:16:27] Sure, sure, we'll do it now, yeah.
[01:16:28] If it doesn't work, if it doesn't work.
[01:16:30] Alternate option also I had given in the past known as Grok.
[01:16:33] Okay, Grok also, same thing again, start building.
[01:16:38] Uh, login.
[01:16:49] Okay, same API keys.
[01:16:51] Create API key.
[01:16:53] Okay, again, same thing, give the name and get the API key.
[01:16:57] Keep Grok also, because in the code, over there, there are both the options.
[01:17:04] Okay.
[01:17:05] There is open router, there is, if it doesn't work, then there is Grok as well.
[01:17:08] Okay, okay, got it.
[01:17:09] If your open router doesn't work, then Glock.
[01:17:14] Okay.
[01:17:15] Huh. Otherwise, guys, you will have to go back to cohere. Cohere is also there. So cohere last week only I showed you using the drag.
[01:17:24] RAG. Non-LangChain RAG we discussed, right?
[01:17:28] Guys, we had discussed it like Streamlit and all.
[01:17:34] Yes, you told us how to…
[01:17:36] you know, build a rapid prototype UI.
[01:17:38] Yes, I'm… Yeah, yeah, yeah.
[01:17:40] And I had also told you that at some point we will do radio as well.
[01:17:46] Okay.
[01:17:51] Okay, let's see how many people are done.
[01:17:54] So I have 10 people done.
[01:17:58] 1, 2, 3…
[01:18:00] What should be the name of this weather API
[01:18:03] Uh, you can give any name, ah.
[01:18:06] Okay
[01:18:07] Vast Deepak.
[01:18:08] Yeah
[01:18:09] And you can give any name, because over here you are
[01:18:12] Deliberately, I'm asking the API key.
[01:18:14] So, over there, you can copy it from there. Let's say I have given weather only.
[01:18:18] Okay, copy from there and paste it.
[01:18:33] Okay, guys, at some point, guys, I would like you all understand GitHub, right? Everybody understand Github, right?
[01:18:40] Over here.
[01:18:41] Yes.
[01:18:42] Yeah.
[01:18:43] Like, I am of the expression that the kind of…
[01:18:49] You know.
[01:18:51] Like, the kind of, uh, place you all are coming from, you all should be aware of GitHub.
[01:18:56] Right? So, at some point guys, I would want, like, you know, what I do in usually other cohorts is we discuss, we do a project.
[01:19:06] Okay. In between, in between only, in between our modules getting completed because see we have been given a lot of time.
[01:19:14] To discuss this agentic part, this LLM, GenAI, this lot of time, okay?
[01:19:19] So there are lots of time we will get to.
[01:19:24] do and explore projects as well, so.
[01:19:26] One point, maybe next week or next to next week.
[01:19:29] I would give you all.
[01:19:32] To execute one project. So basically, let's say I only give you one project myself.
[01:19:37] And you will try to convert that into.
[01:19:42] A modularized piece of code.
[01:19:43] Modularize as in take the code. You can use any coding agent, anything.
[01:19:47] But the main part is, how do you convert them into a modularized code? Bake it down into folders.
[01:19:53] And, you know, put, let's say, ingestion part separately,
[01:19:58] let's say logging part is also there in your code. You have a config Python file from where you take all the configurations.
[01:20:04] Like that, you create a modularized code.
[01:20:06] Okay, it can be a REST API, it can be a normal Python models also.
[01:20:11] No problem, but that kind of a project I would like you all to do because you all have many projects at some point, and instead of directly jumping into the mini project.
[01:20:18] In between also, uh…
[01:20:21] you will have to do. I will give you references, I will give you everything, the project itself only I'll give you.
[01:20:26] The entire code I will give you.
[01:20:28] Your goal will be to convert that into modularized code.
[01:20:31] Okay, so for that, you will need your GitHub ready.
[01:20:35] Because you all will submit things on GitHub.
[01:20:38] Okay, I will talk to Akansha as well, so that your collect, your
[01:20:42] Things could be submitted and
[01:20:45] Later on it could be checked as well so that everybody participates.
[01:20:49] Okay? Easy thing only like I will only give you the project.
[01:20:53] It's already a made up code I will give you. Okay, maybe one of the rack code I will give you. You will just have to convert them into a folder structure.
[01:21:01] Okay.
[01:21:03] Got it guys, fair.
[01:21:10] Yeah, this… yeah, yeah, this.
[01:21:11] Sure. It is… it will be helpful to us to replicate the same thing, what are the project, it will get practical experience even, and we will place
[01:21:18] Yeah, yeah, you see, uh, yeah, Bragaj, just to change your thought process a bit, you will not get
[01:21:23] Exactly enterprise project, but you will get an exposure of how you organize an Enterprise project.
[01:21:30] Got it.
[01:21:35] Yeah, so, uh, see, it's not only how to do the project in a
[01:21:39] Collab because in a, in a collab as in a notebook.
[01:21:42] The point is, how do you organize it, modularize it? That is very important. Okay. In your project submission, also, you have to submit like that.
[01:21:50] Okay, notebooks we use when we start off, usually in a coding journey, many people don't even use notebook. Many people directly use.
[01:21:57] In a modularized way only start off. But I, since I have come from that.
[01:22:01] You know, classical ML deep learning background. So, we use notebook a lot.
[01:22:05] Okay, then we convert this notebook into modularized codes. Okay, nowadays with AI tool, it is very simple only.
[01:22:12] Okay, but you have access to Amazon Q and
[01:22:15] In our enterprise level environment and
[01:22:18] They only give you the folder structure.
[01:22:20] Okay, so in your case, you can give it to ChatGPT and you can give your entire notebook to ChatGPT and ask ChatGPT, like, convert this into modularized code.
[01:22:28] And then you understand that and just do the same folder structure in your
[01:22:34] GitHub and upload it and share the GitHub link.
[01:22:36] And make sure it is running also. And you will have to have a requirements.txt, you will have to have
[01:22:42] Ah, how to start up the project, how anybody can start up the project.
[01:22:45] Okay, so those things also you'll have to mention. You'll have a proper readme file mentioned in your GitHub. So it is most about
[01:22:51] more about making that repository and coding environment setup ready.
[01:22:56] Okay, the project, mostly I will only give you, okay, you just have to convert that. That will be one of the task, we try to do it next week.
[01:23:03] Okay.
[01:23:05] Okay, insufficient grades, never purchased before the key. This is for which?
[01:23:11] Open Router, as I think Azad, you have been using.
[01:23:15] Laguna, right?
[01:23:17] Yes, yes.
[01:23:18] So, are you getting this error for Laguna?
[01:23:21] No, this I'm getting for this one only
[01:23:23] Haha, open router, you have an issue, so you change it to Laguna.
[01:23:26] Yes.
[01:23:28] Okay, okay.
[01:23:29] So, uh, there was a code I had given.
[01:23:31] Uh…
[01:23:38] Yeah, this LangChain open router.
[01:23:41] And over here.
[01:23:50] There was a list of models, no, you can try.
[01:23:52] So yeah, poolside Laguna.
[01:23:54] This one you can try. If your open router is not working.
[01:23:57] Yes, yes, yes. Yeah, I forgot, actually.
[01:24:00] I remember, I remember because your name comes to my mind when I think about Laguna, so that's why I
[01:24:05] Alright.
[01:24:06] I remember because you only discovered that Laguna is working for me. So then I told everybody else. I remember.
[01:24:13] Yeah, yeah.
[01:24:15] Okay, how many people are done, guys, set up ready because you are not left behind.
[01:24:24] 1, 2, 3, 4, 5…
[01:24:34] 7, 8, 9, 10…
[01:24:37] 11, 12… 12 people are done only. Out of 70 people.
[01:24:48] Are you people facing any issue?
[01:24:52] 13.
[01:24:57] Guys, come on, do it. You can't later on say that I got stuck, I got left behind.
[01:25:04] Okay, at least 20, 30 people should be doing it.
[01:25:08] In the class.
[01:25:13] And in the meantime, rest of the people who have done it, you can do your reading, are you reading it guys?
[01:25:18] Are you all understanding people who have done Itazha, then
[01:25:21] Deepak, then Deepan.
[01:25:24] Satish, Aryan, Pawan,
[01:25:28] Sivaraju, Shikhar.
[01:25:30] Vineet, are you all reading it?
[01:25:37] Yes, sir, doing it now.
[01:25:39] Aha, just do it, do it, guys. There is a pyrantic part that you need to understand more.
[01:25:44] Okay, it has to do with that, because pydantix is sim tool calling.
[01:25:47] Format might change later on, okay? Because as you use LangChain, the format will be different.
[01:25:52] But the point is pedantic over here. This is the first exposure you're getting to pydantic probably.
[01:25:57] Okay. Many people don't know even Pydantic. What is what is this? So.
[01:26:02] So just at least read that part.
[01:26:19] Yes.
[01:26:34] Yes.
[01:26:47] This is a plan.
[01:27:09] By the way, guys, how many of you are
[01:27:12] entrepreneurs over here.
[01:27:14] Like fully building your own company or anybody is here.
[01:27:26] Okay, Hemant, I'm sharing it.
[01:27:29] Is anybody there? Like, who is building their own AI startup or some sort of a service-based?
[01:27:35] Thing or a product-based thing.
[01:27:43] Honoratius, tell me.
[01:27:44] So,
[01:27:51] No, I was saying I was trying, or startup AI startup, that's all.
[01:27:57] Oh, okay, ok, so you are in a job, you are doing beside that?
[01:28:02] Oh, Naya.
[01:28:03] Yes, yes, that is a side.
[01:28:09] Yeah, Hayman, I have shared with you.
[01:28:16] So next week, guys, we are also going to a very big event known as Bharat Premier.
[01:28:21] I'll be representing my space startup over there.
[01:28:24] We will be…
[01:28:27] Uh, you know, but we're talking across.
[01:28:33] All the people I think I'll, uh, will be facing a lot of
[01:28:36] Investors come.
[01:28:38] You know, uh, AI influencers, AI.
[01:28:40] You know, these, uh…
[01:28:44] AI people because nowadays any startup event, any expo will have AI people.
[01:28:48] So yeah, let's see what happens. So we are also building.
[01:28:52] Uh…
[01:28:55] You know, ADCS system for space using
[01:29:00] Cheap sensors and, uh, to get the accuracy of a
[01:29:05] One time…
[01:29:08] huge, like, uh, like, you know.
[01:29:11] In, when, when you're…
[01:29:13] uh moving object.
[01:29:16] Moves in orbital.
[01:29:18] It runs through algorithm, so that to keep it correctly in orbit.
[01:29:22] And that algorithm, in order to keep that in a right orbit, you have lots of sensors, probably 100 plus, 200, around 200 sensors.
[01:29:31] Okay, now that depends, like, if it is a satellite, then it will have more sensor. If it is a drone, it will have lesser sensors.
[01:29:37] So, in order to keep it on the right track, you know, it is… it's projectile motions is also correct, okay? It's attitude determination control systems, it is known as ADCS.
[01:29:46] So, we are building our space startup around…
[01:29:49] All right, I know how to use AI and how to use.
[01:29:52] You know, machine learning, plain core machine learning, not your gen AI and all these things.
[01:29:56] But plain core machine learning to.
[01:29:58] Keep that determination in control with lesser human intervention.
[01:30:02] And with cheaper sensors, so…
[01:30:05] Let's say these sensors usually cost a lot.
[01:30:09] Okay, to achieve that kind of an accuracy.
[01:30:12] Getting an ensemble of low-priced sensors.
[01:30:16] .
[01:30:18] Okay.
[01:30:19] who get the accuracy of a higher one sensor. So, instead of taking one big sensor.
[01:30:20] Small, small sensors.
[01:30:21] Yes.
[01:30:22] And we'll try to achieve that.
[01:30:23] Okay, so just go and mute, please.
[01:30:27] Yeah.
[01:30:32] Yeah. So, uh, so that is the thing.
[01:30:36] We are pitching next week.
[01:30:38] So, it is our first…
[01:30:40] Which as well, we have worked on this for one and a half years.
[01:30:43] A lot of clashes we had with mentors and
[01:30:47] Ah.
[01:30:48] situation, sir
[01:30:49] Yeah.
[01:30:52] Which location you are presenting
[01:30:53] Uh, Delhi, Delhi, next week I'll be taking our session from Delhi, so after the event, I
[01:30:59] So…
[01:31:00] Please give that prototype discussion after you successfully completed to us event.
[01:31:03] Yeah, yeah, definitely, uh, at 1 point, so.
[01:31:07] Uh, at 1 point, like when we have more time towards lesser scope, we'll definitely discuss about that.
[01:31:13] Okay, so…
[01:31:16] So, last, uh, one thing I would just tell you, guys, that, uh, before
[01:31:23] 3-4 weeks also we didn't have any idea about what we are going to.
[01:31:26] In a show, this idea was there for last 6 months.
[01:31:29] But we had so many change of mentors, our mentors started, you know, ignoring us after a few days of working, they started ignoring us. So we had a lot of fights with mentors.
[01:31:37] As in, like, they.
[01:31:39] We stopped.
[01:31:41] taking our calls, we gave them equity, everything, but still after that also.
[01:31:45] They stopped ignoring, so we had multiple mentors change, and multiple ideas changed, because of every mentor believed that this is a better idea.
[01:31:51] So we learned from all those mistakes and finally.
[01:31:56] Now, we have a ISROGA in our equity.
[01:31:58] in our cap table as well, so.
[01:32:00] 12 years XRO guy.
[01:32:03] So, he has been helping for the last 6, 7 months. He's not a mentor, he's a technical, pro-technical guy.
[01:32:08] Uh, so…
[01:32:10] He actually helped us along with our guy as well, we built it.
[01:32:15] And, uh, yeah, hopefully something happens. So that's why I was just asking, like, how many of you are
[01:32:20] there in startup journey so that, you know, if somebody is coming to Bharat Premier next week and
[01:32:25] I don't know we can meet as well.
[01:32:28] So that is the thing.
[01:32:30] Anyways, guys, in the meantime, are you all done? Uh, how many of you have
[01:32:36] 1, 2, 3, 4…
[01:32:39] 5, 6, 7, 8, 9.
[01:32:41] 10, 11, 12, 13, 13 people.
[01:32:46] Are rest of you struggling?
[01:32:55] Some might take call from phone, they may not be able to do all these.
[01:32:58] Okay, no problem, no problem, ok, let's start the discussion guys, uh.
[01:33:03] Well, you know, I think this open weather thing, guys, if you all have not done it today, make sure you do it, because our
[01:33:10] Many, any application, we will use this because that is a free open API that we have available.
[01:33:16] order to do the kind of projects that we do.
[01:33:19] In our classes.
[01:33:28] Okay, Vinita, I got it.
[01:33:31] Uh, yeah. Okay, so now let's guys start it. So first we take OpenAI, Pydantic and request all these things. We will be taking request because we'll be making requests to a particular API.
[01:33:42] So we need that. So, first we import that, then we import the open router. Okay, I think these things we have done in the past as well.
[01:33:51] I do know?
[01:33:52] Uh, we create our initialize our models as
[01:33:53] Campbell?
[01:34:00] Hmm.
[01:34:02] Okay, so all these things are there.
[01:34:05] Now, the point.
[01:34:09] Starts, I will go little ahead, ok, because all these things are repetitive, guys.
[01:34:13] The initial basic open router API call, we have done this, like…
[01:34:16] Uh, where we are…
[01:34:25] This is done.
[01:34:27] This is done. Then we are calling the open API. OpenAI.
[01:34:34] Uh, using that same model, and we are passing a temperature of 0.2 and
[01:34:39] A copy of 0.9 and max token, let me reduce this to 150, because…
[01:34:45] This will reduce my token limitation also.
[01:34:50] Okay, so this is done.
[01:34:55] Okay, now guys.
[01:34:57] This still here, guys, we have done before also. Before also, we have done at some and other classes we have done this.
[01:35:04] So, leave that part.
[01:35:07] Uh, so now…
[01:35:08] The point is.
[01:35:11] Look at this now. Now, over here, we are making a call using client.chat.commission.create. Previously, when we were doing it.
[01:35:19] We had a role system.
[01:35:22] We had a role user and there was no such idea about the formatting of the response, how the response will look like response format and all these things.
[01:35:31] Now we have a response format as well.
[01:35:35] So, in the response format, I am mentioning.
[01:35:38] the JSON object that that modality of the response should be JSON object, okay?
[01:35:42] So the moment you respond this, it will not be a plain text answer like this.
[01:35:48] Okay, the ones that you are getting over here, it will be a JSON format answer.
[01:35:59] Like this. Okay, this is the JSON format answer, and we are dumping it on a JSON.
[01:36:03] object, and we are showing the answer.
[01:36:09] Okay, so there is no Pydantic as of now, Shikhar, there is no pedantic as of now. After this, we'll use Pydantic. Now we have just forced.
[01:36:16] The response style to be JSON object, so.
[01:36:18] Ah, I have asked in the system instruction, it returns only a valid JSON object.
[01:36:23] Include the keys and topic explanation, and
[01:36:26] examples. So, if you just write over here, it doesn't matter. If you ride over here, it will be in a JSON dictionary format.
[01:36:32] But it will not be a JSON object. In this case, this becomes a JSON object.
[01:36:36] So, when it becomes a JSON object, you can actually take out only topic.
[01:36:40] You can take, use the key value pair and take out, like how dictionary, how in dictionary you do.
[01:36:45] So that is what JSON object. Why JSON object becomes mandatory? In most of our use cases.
[01:36:50] In most of our practical use cases that we do, guys, we…
[01:36:53] Most of the time it is JSON object guys.
[01:36:55] We hardly take out
[01:36:57] Uh, you know, answers when it is a flat answer that we just want the text answer out of it, then it's fine.
[01:37:03] But in most of the cases, it is, it is a JSON object, guys. Our answers are in JSON only.
[01:37:08] Okay, this text thing is rare, very rare.
[01:37:11] Okay, this takes kind of an output. It is always on JSON only.
[01:37:15] Okay, because many times this JSON object is processed by front-end engineers, other back end.
[01:37:22] you know, components, they might process this and do for other work related to other work as well.
[01:37:26] So, that time it becomes important. So now, this is fine. JSON object, plain JSON object. Now,
[01:37:31] There is something known as a pydantic model.
[01:37:33] Okay. Bidantic model is nothing to do with LLM model. Okay, so by this name never like don't confuse this with model.
[01:37:42] Okay, like LLM model. Pydantic.
[01:37:45] is model in this case replace the model word model with structure.
[01:37:51] Pydantic is more about structure.
[01:37:53] Okay, it makes yours.
[01:37:56] that you follow a particular structure only in order to output certain things.
[01:38:01] Okay, so let's say I create a pedantic model. Over here, Pydantic model I create means I create a class.
[01:38:08] which is inheriting base model. Base model is imported using Pydantix. See, at the top, I have done.
[01:38:15] See, from Biden.
[01:38:19] Some, uh, from Pydantic import base model field validation area error.
[01:38:25] and config dict.
[01:38:27] Okay, these are the things that we are importing.
[01:38:31] Okay, now see, concept explanation.
[01:38:34] is a class that we are inheriting from.
[01:38:37] Base model, uh, base model is again a pedantic from pydantic, so if you
[01:38:42] Want to make a pedantic structure, you create a class like this.
[01:38:45] So, many times in our coding files, we create config files.
[01:38:50] Okay, uh, in our config, we mentioned that how the
[01:38:55] What will be the… which DB from which DB will get the answer. All the DB names.
[01:39:00] All the, you know, ah, configurations, things are mentioned over there. Maybe some env file reading also is done over there.
[01:39:07] Many times people create configurations for your
[01:39:10] LLM output as well. Okay, LLM config.
[01:39:14] People create a file known as LLM config. In that file, people mention pydantic structures like this. Okay.
[01:39:19] Now, this is for one LLM call. You might have multiple LLM calls in different, different LLM calls, you will
[01:39:23] make different, different pedantic structures. So, let's say.
[01:39:26] This is one pedantic structure class you have created also called as pedantic model.
[01:39:31] So whenever you inherit base model and create a class,
[01:39:34] To force a kind of output from your LLM.
[01:39:38] That is known as pydantic model.
[01:39:41] So you are giving this pedantic model you will give in a while, like as you go down, you will see like how I'm getting this as your response format.
[01:39:48] So, over here, you are…
[01:39:50] writing, uh, oh, okay, so model config, you'll have to set this at extra 4 bit some time.
[01:39:56] You know, your LLM outputs extra as well. So you are forcefully making it extra for a bit, like.
[01:40:03] I do not want anything extra. Okay, so please forbid any extra output, because if it comes extra output, it'll give you an error.
[01:40:09] Okay, so now it will not give you an error if you write this, uh, that's why, that's why we have to write this as an extra, you know, statement.
[01:40:17] So that it forcefully, you know, outputs only that.
[01:40:20] Okay, so now comes to the topic. This is the type you are mentioning, that topic has to be SDR. It cannot be a number.
[01:40:26] It cannot be anything. It has to be SDR.
[01:40:28] It cannot be integer. It can be a number, but str number also it can be, ok.
[01:40:35] Like double quotation of 123.
[01:40:36] Okay, but it cannot be…
[01:40:38] Uh, it cannot be an integer.
[01:40:41] dictionary like that. So, topic is a dictionary, uh, is a string.
[01:40:45] And it also describes name of the concept being explained.
[01:40:50] Okay, so this is the topic. Summary.
[01:40:53] Again, string the field, this is also imported from Pydantic.
[01:40:58] This also is imported from Pydantic. Field is imported from Pydantic. A concise explanation in a beginner-friendly language.
[01:41:04] So, your output should have a topic, it should have a summary.
[01:41:08] And it should have a difficulty.
[01:41:11] Difficulty is a collection of beginner, intermediate, or advanced, so it is like a collection.
[01:41:15] Then you have key points. Key points are important ideas the learner should remember.
[01:41:21] Example is 1 practical example, like you want to give and confidence is like a score.
[01:41:27] That is greater than equals to 0, less than equals to 1.
[01:41:31] And confidence score between 0 to 1. Okay.
[01:41:33] So this is your pydantic, and let's just print the pydantic by dumping it into a JSON object.
[01:41:38] This is how the pydantic structure looks like. So, you have made a
[01:41:42] This extra for bid.
[01:41:44] This also means additional properties equals to false. So if you mention this.
[01:41:49] Okay, many people do this, many people does this also. Additional properties equals to false also, many people mention.
[01:41:54] Okay, but this is the same thing. Like if you make extra for bit.
[01:41:59] then this automatically becomes false.
[01:42:01] Okay, so any additional things will not be there, okay? So, this is how the formatting
[01:42:06] The structure output is expected, ok.
[01:42:09] And, uh, so you have, uh, made sure that.
[01:42:16] Yeah, so basically in summarization, required.
[01:42:19] uh, key-value pair, uh, keys are topic, summary, difficulty, key points, example, and confidence.
[01:42:25] Okay, concept explanation is the model.
[01:42:28] Pydantic model name.
[01:42:30] Okay.
[01:42:32] Now, let's go down and let's see how an output will look like.
[01:42:37] Okay. You are an AI instructor. Follow the supplied JSON schema exactly. This is the system instruction.
[01:42:44] And this is the dynamic user instruction that you can place. Explain LoRa scaling.
[01:42:50] In simple, simple language.
[01:42:52] Okay. And you pass the JSON
[01:42:57] Schema over here, ok, you write type equals to previously you were writing JSON object, now you pass JSON schema.
[01:43:04] And what is the JSON schema? You write that. You write the pyrantic, you give a name to it.
[01:43:09] You give strict that you will have to follow it.
[01:43:11] And and the schema, followed by that pyrantic model class, .model JSON schema.
[01:43:16] Like this you pass, then automatically this entire thing goes to your LLM.
[01:43:21] Your LLM able to see this entire thing, or you can say it is able to see this entire thing.
[01:43:26] This, this thing.
[01:43:28] Okay, model JSON schema now.
[01:43:30] It is able to see this entire thing, so.
[01:43:33] Along with this, along with this, this also goes to your LLM.
[01:43:38] This also goes to your LLM.
[01:43:40] So your LLM forced to output in this format.
[01:43:43] Okay, so your pedantic.
[01:43:47] Is kind of like a thing which actually forces your LLM to output in a particular format.
[01:43:54] Okay, now, once you get the output,
[01:43:57] You can, you know, validate the output as well.
[01:44:00] Okay, it will show like this, you know, response validated pydantic validation.
[01:44:06] Okay, and it will show you the response. That is how it is showing. So, there is this validated JSON.
[01:44:11] Okay, model.validate.json. If you do and pass that response that you have.
[01:44:15] through it, if everything is fine, then you will not get an error. Otherwise, you will get
[01:44:19] a validation error. So validation error is also imported at the top if you see.
[01:44:25] See, validation error.
[01:44:28] So, Pydantic is mostly done for this thing. It is used to structurize the input and the output. Mostly people use it for output only.
[01:44:36] Rarely people use it for input, but input also people you can use it for.
[01:44:40] Okay, mostly people use it for output because input is mostly user.
[01:44:45] uh, guided, na. So, user might not follow this format, okay? User might give an error or something, so that's why user input is not so much followed, but output it is.
[01:44:55] Definitely followed here, Zaditya.
[01:44:58] So, when you say output, it is LLM…
[01:45:01] structuring the…
[01:45:03] such and the information in certain formats, so that the tools can accept the request, right?
[01:45:17] Okay. Hmm. And…
[01:45:18] Got it, Aditya, so let's say after this output, I take output, let's say I take like this, for example.
[01:45:25] Uh, what was your output?
[01:45:27] So let's say after this output, I require the topic out of it.
[01:45:32] Okay.
[01:45:33] For further related work, okay, any other work.
[01:45:35] Okay. Okay.
[01:45:36] Okay, so…
[01:45:38] So if it is structured like this, then only I can take it. If it is not structured, let's say if this is, instead of a JSON, it is a text.
[01:45:44] But in this format only text, but in our dictionary format.
[01:45:47] Okay, then if you just put a key like this, nothing, no answer will come. It is a text, it is a plain text.
[01:45:52] Yeah, okay. Hmm.
[01:45:53] Got it. So, that is why, uh, you know, pyrantic.
[01:45:57] is used so that your output is, first of all, uh, first of all, you get a structured output.
[01:46:03] And it is a validated output also. People will not start giving rubbish things.
[01:46:08] Okay.
[01:46:09] Your LLM will not start giving rubbish things because sometimes your LLM might
[01:46:12] hallucinate. If you do not give this format.
[01:46:16] Hmm.
[01:46:17] Okay, so this, like, it goes like a prompt engineering only, it is like part of your prompt engineering.
[01:46:22] Okay.
[01:46:23] Yes.
[01:46:27] Yes, yes, please go ahead, who had more question, yes, Pallavi.
[01:46:35] Yeah, so, could you please explain, the model config part once more? Like, config tech thing.
[01:46:43] Oh, config dig thing, this one.
[01:46:45] Yeah.
[01:46:46] Uh, so his input also in the same format, or is it, like, any input is, um…
[01:46:47] And lower than the response only is generated in the JSON format.
[01:46:48] Oh, this one is, see, sometime, if you don't give this, na, your
[01:46:49] Your LLM is in a dilemma that whether it should output extra things or not, additional properties or not.
[01:46:55] Other than these properties.
[01:46:57] Got it? Your LLM is in doubt.
[01:47:02] Okay.
[01:47:03] Okay, that's why we write this x equals to 4 bit because if I don't write this, you will get an error over here. When you make the LM column.
[01:47:07] You'll get an error about additional properties.
[01:47:11] Okay, so it is waiting that
[01:47:13] Okay.
[01:47:14] I will give this, but shall I give more additional properties?
[01:47:16] So, you are just writing extra equals to forbid, no, no, don't give anything else.
[01:47:24] That's all.
[01:47:27] Uh-huh.
[01:47:31] Okay, so now guys, uh, now.
[01:47:34] Let us go ahead.
[01:47:39] And to, uh, use our open, open weather.
[01:47:42] the way we have taken it. So, first of all, your open weather, you will be, you know, asked to provide the API key.
[01:47:48] Hopefully, guys, it works for all of you. Like, see, we are at the like at the.
[01:47:53] And the hope of the open air only app only, like, I don't know, like, it might not work for
[01:47:58] All of you, if it is not working, then you'll have to sign up with a different email ID.
[01:48:03] Alright, just a moment.
[01:48:36] Yeah, okay. Now we take the weather app. I hope, guys, everybody is able to use. There are people.
[01:48:42] Who has, like, you know, that is.
[01:48:44] Due to open weather, like, I think they also have that lottery system, like open router. Everybody won the access.
[01:48:49] Because this I have recently noticed,
[01:48:52] Okay, in most of the cohort, it worked. Recently, I have seen like one or two students, it is failing.
[01:48:57] Okay, if it fails, you just try with a different account, okay?
[01:49:03] Now guys, so this is open weather, this is your open weather function call, so open weather function call has a URL.
[01:49:11] uh, has…
[01:49:12] uh no request.get.
[01:49:14] uh, you have to pass a location, you have to pass the API id, the key.
[01:49:20] API key, and the units, and there are more metrics that is there in open weather.
[01:49:25] Okay, and you have to, you can mention a timeout, like, like 20 seconds timeout.
[01:49:29] Okay, I think this is 20 milliseconds for them. Okay. And the response will come in this format, location, temperature, humidity, wind speed, and condition.
[01:49:37] Okay, this is a very standard format of open weather.
[01:49:40] Okay, so your API might have a different format. Let's say the kind of API you will be using for your enterprise will also have a similar kind of a JSON structure that is being returned.
[01:49:49] Okay, so imagine, like, instead of this, you will have your company APIs.
[01:49:54] Okay, now just test it once.
[01:49:57] If you just press this, this is the option that you will get.
[01:50:00] Okay, if you just search with location, see Kolkata temperature.
[01:50:04] immunity, all these things are coming. Fine.
[01:50:07] Fine, this is coming. Now guys, the point is, many people.
[01:50:11] The point is, your question is
[01:50:14] That when I'm doing this chat,
[01:50:17] Many people will not write Kolkata.
[01:50:19] Many people will not write Mumbai, many people will not write Delhi.
[01:50:22] They will write what is the temperature of Kolkata today.
[01:50:25] They might write like this.
[01:50:27] What is the temperature of Kolkata today?
[01:50:29] So that is why void, what do you need over there?
[01:50:33] What can solve that problem?
[01:50:37] What do you think can solve that problem?
[01:50:43] Entity recognition
[01:50:45] Um, okay.
[01:50:47] Understanding the place and then keep that
[01:50:48] 15. Karampani pillow.
[01:50:49] Fine, fine. But just now we learned it.
[01:50:52] LLM to identify location.
[01:50:55] Yeah. But just now we learned it Pydantic.
[01:50:59] So, you can use Pydantic in your input also. So, this is the format. So we are creating a Pydantic model for input now.
[01:51:05] Same thing, again, you write again pyrantic, oh, this is not required, we have already imported at the top.
[01:51:10] So, class weather tool input.
[01:51:14] Pydantic again, base model.
[01:51:15] Okay, location, if…
[01:51:18] the name of the city, for example, Kolkata or Paris. Whatever you give.
[01:51:21] If you pass this to
[01:51:23] this pedantic input, you know, schema, through this schema, when you are making that LLM call.
[01:51:28] It will just take out the city from there, because this will also be accessed by your LLM. Your LLM will get this structure. So your LLM is kind of doing entity recognition only, Nipan.
[01:51:38] Whatever you have told, you have told correctly that the thing is that.
[01:51:42] Pydantic is going to solve that.
[01:51:44] Correct.
[01:51:45] Okay, so now guys, we are creating a plain dictionary. Usually tool callings.
[01:51:50] Tool calling before LangChain, without LangChain, if you ever do it, it is done using dictionary, very standard format.
[01:51:56] a dictionary where you mention type.
[01:51:58] function is get, uh, after function, you write the name get current weather.
[01:52:04] Uh, then description of the function, all these things are accessed by your LLM. Your LLM is going to see this.
[01:52:10] So the more defined you write, the better it is.
[01:52:13] And the parameters, the parameters will be weather tool input models JSON schema, so this input it has to follow.
[01:52:21] Okay, so you are passing this.
[01:52:23] Okay, so see now if you look at…
[01:52:29] Yeah, if you look at this parameters under parameters, can you see a
[01:52:33] The entire thing got evaluated, elaborated.
[01:52:36] Over here, you just wrote parameters whether tool input model.
[01:52:40] JSON schema, automatically see the description, name of the city,
[01:52:44] title, location, string, everything came in.
[01:52:46] Okay, so this is going to be accessed by your LLM. Your LLM is going to see this.
[01:52:52] Directly, I'm going to see this, so your LLM automatically, if you mention
[01:52:57] What is the temperature of Kolgatta? It will extract that Kolkata out of it.
[01:53:01] Okay, from the input, and then it will be passed to the weather API.
[01:53:05] Okay, now you create a list like this, a list in a form of a dictionary only.
[01:53:11] Where you write all the available tools you have.
[01:53:14] So we have getCurrentWeather. If you have more tools, you will write more tools over here.
[01:53:18] Okay, so when we do agentic, I will show you, like, how you can you can have more tools as well.
[01:53:23] So, this is like, you know, list of tools you write. You can have, let's say,
[01:53:27] Get, uh.
[01:53:30] Get weather API, get, uh, let's say, uh, for example, get search responses.
[01:53:35] Okay, get mathematical expressions being solved, so.
[01:53:40] Different, different tools might be there. Okay, so over here, we have one tool.
[01:53:44] And we are checking.
[01:53:48] We have created an execute tool option. What execute tool option will do?
[01:53:53] It will check.
[01:53:54] If tool name is mentioned in that available tool or not.
[01:53:58] If, from the available tool, the tool name is not mentioned, then you will just say unknown tool requested.
[01:54:03] If it is there, if the tool name equals request to getCurrentWeather,
[01:54:06] Then please, from the weather tool input,
[01:54:09] Validate JSON, just like we have done validation over here, guys.
[01:54:14] We had done validation over here, validate JSON.
[01:54:16] From there, you validate the JSON. So first you validate the JSON.
[01:54:20] You know, using that pydantic model, you validate the JSON and validate the output.
[01:54:25] If the output is fine.
[01:54:27] Then using that available tool calls and the key name, key name is getCurrentWeather.
[01:54:31] Pass that parameter.
[01:54:34] Okay, so inside this you pass that parameter.
[01:54:37] So this will automatically call the getCurrentWeather, ok, so just play this.
[01:54:43] Now, let's look at the entire tool calling workflow.
[01:54:46] Okay.
[01:54:48] See, with this you will understand everything. So first,
[01:54:51] Look at the flow. First, we have an LLM call.
[01:54:54] Okay, in the LLM call,
[01:54:57] We have the model name, same messages, and we have a tools option, tools parameter.
[01:55:03] So, when you make client.chat.commission.create.
[01:55:07] Till now, you have been doing till here.
[01:55:08] Okay, and here, these two have never seen.
[01:55:11] Okay, so there is something known as tools.
[01:55:16] In tools, you can literally mention all the tool definition like this, whether
[01:55:20] With commas, you can give search.
[01:55:22] tool definition like this.
[01:55:26] Like this, you can give all the tool like this, all the tools that you will ever mention, you will create.
[01:55:30] So you can mention all the tools.
[01:55:33] And automatically that tool definition that you have, this will go to your LLM.
[01:55:38] This entire thing goes to your LLM, so your LLM understands. So not only this is not only your prompt, your prompt includes this as well, the tool access, the tools list of tools as well.
[01:55:47] Tool choice is auto, so there are other options like strict as well, so then it becomes forcefully to use that.
[01:55:53] Okay, uh, that tool only, but it is auto, like, based on the LLM's brain.
[01:55:58] LRM only take the addition of using that tool.
[01:56:01] Okay, so we are not forcing anything.
[01:56:03] Okay, so usually, uh…
[01:56:06] Let's keep it auto only usually in agentic, we keep it auto. If you want.
[01:56:10] To have a predictable pipeline, then we have LangGraph over there, we actually
[01:56:14] you know, make a predictable pipeline, let's say, I want this tool only to be used from there.
[01:56:19] this tool like that. We have a predictable pipeline over there.
[01:56:23] Okay, so this is your tool, this is your temperature.
[01:56:27] And temperature you can change, you can make it 0.1, 0.2, but usually when you have a factual information coming out.
[01:56:35] Uh, when you have a particular structure to be followed, usually we take a temperature of 0.1, point two, point three max.
[01:56:41] Okay, so whatever we get the response, we get a response, ok.
[01:56:45] We get a response that we stored in assistant message.
[01:56:48] Now, we check.
[01:56:50] If in the assistant message there is no tool called.
[01:56:54] If in the assistant message there is no tool call.
[01:56:57] Then just return the content.
[01:57:00] So there could be no tool call as well. Let's say I might ask, like,
[01:57:03] Uh, why does…
[01:57:05] Uh, you know, humidity, uh, there is one question I had asked.
[01:57:09] Why does humidity sometimes make hot weather feel uncomfortable?
[01:57:14] So this doesn't have any Kolkata or any city mentioned.
[01:57:17] So directly the answer you can show.
[01:57:19] But if there is a tool called
[01:57:21] If this is not the case, there is a tool called
[01:57:24] Then, you know, catch that tool call first, get the tool call.
[01:57:29] And the argument of the tool call, you will also get, have the argument that is being returned. Toolcall.function.arguments if you do.
[01:57:35] You'll get the arguments as well.
[01:57:37] Okay, and…
[01:57:40] You just call this getCurrentWeather.
[01:57:43] Weather requires to get current weather.
[01:57:45] So, get current weather, pass that argument location.
[01:57:49] Automatically, it will call the get current weather.
[01:57:52] Which you have created over here.
[01:57:54] get current weather. You have created.
[01:57:57] And with that,
[01:58:00] Your thing is done and then whatever is the response that you get, you get a response after this.
[01:58:07] Okay, after this you get a response, weather result.
[01:58:10] Now that weather result, you append it, add the tool call and the result to the message.
[01:58:15] Okay, and then
[01:58:18] Again, again, you do another LLM call to beautify that message.
[01:58:23] Okay, to beautify that message.
[01:58:26] So let's do this.
[01:58:35] Now this is done. Now guys, if you ask this, see.
[01:58:38] What is the current weather of Kolkata?
[01:58:41] GetCurrentWeather, see, this is the argument that you got. GetCurrentWeather is a tool called that all these things are getting printed, guys.
[01:58:47] tool called current weather, all these things.
[01:58:52] Okay, uh, then weather result.
[01:58:55] Uh, you have location Kolkata.
[01:58:58] All these things, this is the entire weather result.
[01:59:00] Then you have a second LLM call, from where the beautified answer is coming in a markdown format, it is coming.
[01:59:08] Okay, so this is the second LLM call answer. Now, if you ask a question like this, why does humidity
[01:59:14] Sometimes make hot weather.
[01:59:16] Feel more uncomfortable, then you do not have any, any tool called or anything. Directly you get the returned answer from here.
[01:59:26] Return answer from here.
[01:59:29] Got it?
[01:59:33] Everybody understood tool call, now people, first timers.
[01:59:38] Did you all understand tool call?
[01:59:44] First timers, guys, do you think this is very complicated? See, my purpose was not tool call over here. Tool call, again, you will see another example.
[01:59:51] My purposes, did you understand pedantic?
[01:59:56] So basically, it is about how a tool is being called, is it? That is what it is if I just put it in simple…
[02:00:02] Okay.
[02:00:03] It is… our tool will… yes, tool is being called and how do you ensure using Pydantic that you follow a particular structure.
[02:00:07] No, good.
[02:00:08] And that pyrantic actually determines all your input, your output, and all these things.
[02:00:13] Got it.
[02:00:14] So, Pydentic, if it's… so it primarily used for structuring everything, is it? The… the output…
[02:00:18] Yes, it is, there is no other reason only to use it.
[02:00:21] work okay. And it is used everywhere, uh… an urban, or… it's like, I don't know, a standard that…
[02:00:28] Everyone uses, or is there any alternative to Pydentic which people might be using?
[02:00:31] No, no, no, there are, uh, no, no, no, either you will use Fireantic or you will not use at all. There is no alternate to Fireantic.
[02:00:38] Okay, you can use the
[02:00:44] Yes. Yes.
[02:00:45] Sir, I have one more question, sir. I have seen Pydentic is more like a validation to the output and styling. And, if it… see, I have not seen anything like, we need to call out Pydentix
[02:00:54] In the code. It is definitely if we control the structure output and syntax, and in runtime errors will be controlled by this way, that's all, correct? Nothing especially to call it.
[02:01:04] Yeah, Pydantic, you basically use that so that your output is structured because
[02:01:10] Your output might be an input to another LLM or another agent also.
[02:01:15] So that time you force it to use it. That's all. If you do not use Pydantic, then you can.
[02:01:21] Make it a normal JSON structure as well, like normal dictionary, you can mention a dictionary that this is the format you should follow.
[02:01:26] But Pydantic, you will see it becomes a validator.
[02:01:30] Okay, as a validation framework. Got it, Vineet. So that
[02:01:33] You know, your output is always validated.
[02:01:35] Okay. So, so…
[02:01:36] Okay, sometimes your outputs are not within the limit, it might give you greater, greater than 1 value.
[02:01:42] Got it.
[02:01:43] So, how do you make sure? That is how you write Pyranty. Pydanti is had made our life easier, that.
[02:01:48] Okay, so other alternatives, Anirban? Like, uh, or rather, where is Pydentic used more, or is there an alternative that people have started using?
[02:01:55] Uh, no.
[02:02:00] Okay.
[02:02:01] No, Pydantic has been followed before LLMs also. So guys, you know that Pydantix had been used before LLM as well. It got the popularity before.
[02:02:04] With LLM, okay, it was used even without LLM also. So, that is, uh, that is why it is, it is used by ranting even on normal.
[02:02:15] Okay.
[02:02:16] You know, in API responses also we use Pydantic. Like, I want response model, like I want my.
[02:02:18] A responses to look like this only.
[02:02:20] Okay, so that from my API, I get response in this format only or let's see if I'm building an API.
[02:02:25] I get this format only. So Pydantic was used before also its
[02:02:29] It's just like a standard.
[02:02:31] uh, validation framework.
[02:02:35] Mm-hmm.
[02:02:36] Got it. It's like a industry standard, kind of, if I put it in very simple terms. Okay.
[02:02:37] It is not like it is not.
[02:02:40] Uh, like LangChain that it is something externally created.
[02:02:45] Okay.
[02:02:46] It is something, like, within the Python community, only it is used.
[02:02:49] Okay.
[02:02:50] So it is not externally created that you have alternate, like.
[02:02:52] Alternative to LangChain is Llama Index, for example.
[02:02:54] Okay.
[02:02:55] Okay, but it didn't kick off, but that's a different story. But yeah, that like those kind of things has alternate.
[02:03:00] Okay, but pyrantic is…
[02:03:02] It's like, it has become a basic thing now for using anything.
[02:03:07] Okay, got it.
[02:03:11] Okay, so guys, uh, so this is what tool calling is, ah, next day, tomorrow we will do about agents, uh,
[02:03:19] Language and agents, where we will use multiple tool calls, not only this.
[02:03:24] Okay.
[02:03:25] Sir, what time tomorrow it same 5 o'clock or 3 o'clock, sir? I got 2 invitations
[02:03:28] I don't know why only we are sticking to 5 till August.
[02:03:32] Okay, sir.
[02:03:33] Okay, because few people have requested, uh, time change, so August, at some point, uh, my
[02:03:38] Uh…
[02:03:39] It's fine, sir, 5 o'clock is fine, yeah.
[02:03:41] Yeah, my… no, only this for, again, for everybody as well, uh, let me tell you.
[02:03:47] Just a moment…
[02:03:49] Just, uh, I'll clear the timing today only.
[02:03:55] Okay, so there is a
[02:04:02] There could be a time change.
[02:04:05] From, yeah, August only.
[02:04:08] From around 20… mid of August, 22nd or 23rd.
[02:04:14] So from there onwards, we will change the time.
[02:04:17] Okay, that's all.
[02:04:20] Okay, I'm in discussion, so we will change the time from there. Then from there till September.
[02:04:25] Wherever September, mid, or September end, we will…
[02:04:28] Follow that time.
[02:04:31] Okay.
[02:04:33] Okay, guys, uh, thank you guys. I will, you know, wrap it here.
[02:04:38] uh, wrap it up here. Tomorrow, I think we will have, ah, agents discussion.
[02:04:44] And, uh, let's see, uh…
[02:04:47] What turns out, and then we will also.
[02:04:50] Give a little bit of introduction about
[02:04:51] Blan graph as well.
[02:04:55] Just one question from GenAI Batch Manager. The thing, is
[02:05:00] Akansha, are you there?
[02:05:03] Yeah.
[02:05:04] Wait.
[02:05:07] Yes, yes, sir, I'm here
[02:05:09] Yes, yes.
[02:05:12] Yeah, so the thing is, apparently there is one more, discussion, or, you know, this kind of session from Ashutosh
[02:05:22] So, couple of us, we did not get an invite for that
[02:05:26] Could you please repeat for what you… you're asking for?
[02:05:34] Okay.
[02:05:35] There's one more session, I think it happened on 3 o'clock today, and a couple of us, we did not get an invite for that.
[02:05:38] Okay, for AIOps got 3
[02:05:42] Yeah, it is AIOps, right?
[02:05:45] Okay, okay, let me update.
[02:05:48] Okay, Ashuto server there, right?
[02:05:51] Okay, okay. Okay.
[02:05:52] Yeah. Along with me, there will be a couple of other folks who will not get the invite.
[02:05:57] So probably they can also
[02:05:58] Okay, I'll ask them, okay, sure.
[02:06:01] Thank you.
[02:06:02] Yeah.
[02:06:05] Yes
[02:06:06] Now, Akansha, one more question. So, the pathway to success-related sessions, are these also recorded and then kept it in the portal for looking at it later?
[02:06:10] Yes, sir. Yes, sir. They are recorded, and they are uploaded on LMS
[02:06:15] Okay, I don't see it. If you can
[02:06:19] guide me where exactly I can look at it, be great.
[02:06:22] Sir, I'll forward that particular recordings to you.
[02:06:26] Okay. You know my name, right? Deepan.
[02:06:27] Okay. Yes, yes, yes.
[02:06:30] Okay.
[02:06:31] Yeah. And I think just one more… just one more thing. So, in event detail, we usually getthe details for the regular session, but anything extra, we don't, see that over there on the… on the
[02:06:43] So, you aren't enrolled on manager pathway?
[02:06:47] LMS part, you are enrolled there, the notification is there, announcements are there.
[02:06:54] On the portal?
[02:06:55] Yeah, yeah, there's some question for you from Deepak.
[02:06:56] Yes, on the NMS portal, announcements to Hoti and Abrah.
[02:07:00] Okay, at Daresh also, by the way, Akansha. Nadesha has messaged me, at Naresh also.
[02:07:01] Yes.
[02:07:05] That is, you can message directly Akansha.
[02:07:08] GenAI Batch Manager.
[02:07:14] And…
[02:07:15] I think there's always a confusion with the LMS portal. I don't know, it's not very
[02:07:17] easy to navigate to find things.
[02:07:25] Correct, yeah
[02:07:26] Actually, there were a couple of total, but I think now the current one is IATR featurents.com, right?
[02:07:27] No, but I think for the pathways, we have the pathway success also, we are to see for…
[02:07:32] different protocols, it doesn't comes into…
[02:07:35] The new portal, so, I mean…
[02:07:37] Then share the line of this group over there.
[02:07:38] But whatsoever, whatsoever is announced on Pathway to Success, for that, you get email also, no?
[02:07:45] Don't you guys get email
[02:07:48] So, we get email, but for those recordings, I think, see, typically, even the next session is going to be 7.30. If we again join 7.30 session at home, wife will kill less for sure. So, we prefer doing a recording or viewing the recording
[02:08:03] But we don't get any, anywhere that information. Some of the sessions are also really good. So we're missing out all of them right now from a recording standpoint.
[02:08:16] Okay.
[02:08:18] And also, Kansai cannot see any video recording for this Module 5 AI Agent, and…
[02:08:24] Agentic frameworks, and even I cannot see any official material also.
[02:08:29] So, heavy…
[02:08:30] Okay, I'll upload that online.
[02:08:33] Okay.
[02:08:35] Your last few notebooks are missing there.
[02:08:41] Okay, we'll see, we'll see today.
[02:08:46] Yeah, in announcement, I could see this mention of these sessions, but nowhere is there in the event. Also, I'm not sure how to get the recordings of these. I could see some code session, GenAI manager cohort, and some session on the code update
[02:09:01] Something like that, but I don't see the events
[02:09:05] Let me check all the recordings today.
[02:09:08] Yeah.
[02:09:09] And I'll upload whatsoever is missing there. No issues.
[02:09:11] Also, ensure that we get the invite, because I've checked the email, I'm not getting the invite for that. I get it for regular sessions, but not for that.
[02:09:21] Okay.
[02:09:22] Uh, please work on this reference material also, that is very important while we are going through this reference, right? We are not getting any reference material, what Anir is talking, uh, teaching, right? So…
[02:09:31] That is needed for each and every video to be linked.
[02:09:37] Okay, it should be there, I understand. I'll upload.
[02:09:40] Please, please, yeah.
[02:09:44] Sure, sure, no issues.
[02:09:45] Maybe, for next session, right? Next session, as in next week's session, before even we start with it, if you can provide some updates on all of our questions right before we get started with the session, that would be great
[02:09:56] You take 10 minutes and then tell us what else you have fixed from next Saturday. Again, you take the whole week, that's fine.
[02:10:05] But if you can clarify all those questions, what we talked about today in the first 10 min. That would be great.
[02:10:13] Yeah, also, can you summarize about what all meetings we have tomorrow, like, on Sunday, so that we don't miss out on anything?
[02:10:23] Well, you guys have got the notifications related to the tomorrow's classes now?
[02:10:29] Yeah, so on the portal, I do only see session from 5 to 7, but I've got email for different session as well.
[02:10:38] But I'm not sure if I'm missing out on anything.
[02:10:44] Sorry, I didn't get you.
[02:10:48] So, what I'm saying is, on the portal, I only see the notification for 5 to 7, this regular session, but apparently, I've got email invite for other session
[02:11:00] But how do I ensure that, you know, I'm not missing out on anything?
[02:11:03] Then how did you join this session?
[02:11:09] This regular session, I've got the invite, as well as
[02:11:11] You are saying… you are saying about career assistance and all, you didn't get any notification, right?
[02:11:17] Yeah
[02:11:19] How is it possible for the students at working from? Okay, let me see, like, might you are left for registration there. Let me see…
[02:11:33] Meanwhile, if I have to email you, if this is not resolved
[02:11:36] Yes, you can. That works.
[02:11:39] Can you pick the email?
[02:11:46] Sir, can you message the email, yeah?
[02:12:16] You can write email to the career support, that works.
[02:12:22] Okay, sure.
[02:12:23] Okay, okay.