Відео

I am glad that it was helpful!

Thanks! I haven't tested it in that scenario but I think it should work. As long as the DB that is getting updated is the one that is passed to the agent, the agent should be able to interact with it in real-time.

I am glad the contents were helpful!

Yes you can use local LLMs. But the LLM must be very powerful and able to handle long context lengths. Plus, the agent itself needs to be adjusted and modified to work with the O.S LLM

I didn't understand the +-60% accuracy. But if you are referring to the accuracy of this technique, I should say this approach is probably on the frontline of retrieval techniques and it is very new. So,I expect it to become better and better over time with the advancements in the field. I also should say that between GraphRAG with tabular data and using Graph Agents to directly query the DB, I prefer the second approach and I would use agents

Thanks. I am glad you liked the video. I removed the function calling video because OpenAI updated the functions and they changed the way the models were called. But you can have access to the code here: - github.com/Farzad-R/LLM-Zero-to-Hundred/tree/master/tutorials/LLM-function-calling-tutorial plus, I am using function calling in the following two projects: - ua-cam.com/video/KoWjy5PZdX0/v-deo.htmlsi=DN1Gt6sA8W-E4C2l - ua-cam.com/video/55bztmEzAYU/v-deo.htmlsi=kMV5ZtPaGugVckP6 And in the next video I will explain the new ways of function calling and how to design LLM agents from scratch. That project is almost ready!

Thanks!

This system works on MacOS. If the provided requirements.txt file doesn't run, install the necessary libraries manually. Once you have them, you're good to go.

@@airoundtable Thank you so much! I have a question regarding RAG: I am currently working on a survey data with open-ended questions and mostly user type-in answers (which means they are unstructured and may vary by personal writing/spelling/Abbreviation habits, like ai/AI/A.I, etc), and I wonder what approach would be the best for me to analyze and do Q&A with the data, especially grouping and pairing users by their answers? I know sql query is no use due to the nature of unstructured responses, and traditional RAG that don't use the model's own knowledge but solely the context (I remember seeing this in one of your video, correct me if I am wrong...) may not be helpful because I need the model to read these answers like a human not a string matching machine.

@@黄迪宏 You're welcome! I need to know more about the source of the data. When users ask their question, what type of a data do you want to search for getting the answer? Is it tabular or semi-unstructured or fully unstructured data?

@@airoundtable Its an excel file of google survey results, each column is the answer to one question and each row refers to one respondent. Most of the questions require typing instead of selecting a option, so the answers will be all unstructured strings. I would like to have the model answer my questions like who and who have similar interests in what (grouping), whose skills seems to be able to solve whose problem (pairing), who and who know the same people, etc.

@@黄迪宏 Interesting.. I never worked on such problem. But to quickly brainstorm here a bit, if the type of questions that you want to ask from the DB are clear and you know what questions you are looking for, you can use an LLM and construct a knowledge graph from unstructured data and use a GraphAgent to query the graph DB that you created using the knowledge graph (I have a video on knowledge graph and in the final part, I explain a similar problem for a medical chatbot). With this approach you can at least answer all the questions that you mentioned in your previous reply. However, if the type of questions are not clear and it is an open ended problem and users can ask literally anything from the DB, that would make it a much harder problem to solve. Because not only you want to search the text but you also want to connect the dots between multiple possible answers and make a conclusion from them. I am not sure what would work for that scenario at the moment

Thanks. I am glad you liked the video. yes, I have a video explaining how to design a RAG chatbot using Open Source LLMs, which might interest you. here is the link ua-cam.com/video/6dyz2M_UWLw/v-deo.htmlsi=XvWfkFAEL4Jt4CsX

Thanks. Add agents specifically designed for data visualization to the system. That would be my first choice since current agents are very good a retrieving data but they just was not desgined to visualize the data.

Thanks! I am glad you liked the video

Thanks for the suggestion. I will try to include more applications with O.S LLMs.

That is a great idea. Adding few shot examples and extra content can improve the performance

@@airoundtable Today, we achieved excellent results by incorporating a few-shot learning approach, enhancing them with dynamic vectors, so eventually dynamic few-shots. We also developed a concise table description aligned with the few-shot prompts. These results are highly optimized, significantly reducing calls to OpenAI and token usage. Even with GPT-3.5-turbo, we've used fewer than 3,000 tokens so far.

@@debarghyadasgupta1931 It is great to hear it! I really like the strategy and it is very interesting to me that the number of calls to OpenAi were reduced as well. Great job and thanks for sharing the results with me!

@@debarghyadasgupta1931 Hi, good to hear your great progress! Would you mind sharing some insights about how to do it? I am interested in how to incorporate few-shot and dynamic vectors.

@@黄迪宏 The few-shot examples were optimized using vectors. First, you need to embed all your few-shot examples. Then, when a user poses a question, you perform a vector similarity search exclusively on these few shots. Typically, Approximate Nearest Neighbors (ANN) is used for this purpose. I prefer using k=5, so among the 100 or possibly 1000 few-shot examples in your collection, you will identify the top 5 that are most relevant. Adding context from the table can further refine the query. This approach not only reduces the number of calls to the LLM API but also decreases the token usage.

This is the abstract of my project:*"An Intelligent Chatbot for Attendance & Academic Support in Educa􀆟on" This study introduces an AI-driven chatbot designed for educational institutions, focusing on attendance management and academic support. Leveraging natural language processing (NLP) and machine learning (ML), the chatbot simplifies attendance tracking for educators and provides personalized academic guidance for students. It offers real-time updates on attendance records, performance metrics, and suggests tailored improvements. Additionally, serving as a communication and data visualization platform, it offers dynamic representations of statistics, enhancing decisionmaking. Overall, this intelligent chatbot streamlines administrative tasks and fosters student success in education

Thanks! OpenAI models are not free. But you can try to build this project using opensource models from huggingface. Search langchain agents using huggingface models and you would fine some guides on how design the agents using opensource LLMs

It is an interesting project. With a good database, you can design an agent to accomplish the goals that you mentioned in the project description.

​@@airoundtableyeah thanks , many discouraged me that I am studying 2nd year in 3rd tier btech college you couldn't do it , but I am willing to and will see where it goes , Thanks again 🙂.

​@@LIVELOVEASH645 Don't let anyone discourage you. If you think you can, you can.

It can perform simple statistical analysis like mean, median, standard deviation, etc. I am not sure what you meant by "answer questions about the data in a comprehensive and informative way". But since the agent has access to the database schema and can also be provided with further information such as examples and details of the DB, it can have a very good understanding of the DB and would be able to answer general questions as well. It cannot plot the data and for that, you need to improve the agent's capabilities and possibly add a new agent designed for visualization tasks to the system.

Firstly, it's important to note that selecting strategies based solely on token usage isn't practical, as each strategy serves a different purpose. Generally, over the long term, LLM agents tend to consume more tokens compared to standard RAG approaches. This is particularly true when multiple agents collaborate or require frequent debugging. In RAG and similar approaches, the primary cost arises during initial data vectorization. After that, it functions as a straightforward chatbot with lower token consumption compared to agents.

Hi Harry. Most probably you asked it on the right video :)). In general, the approach that can work best very much depends on the type of questions that you would like to ask from the database. As a simple rule, if your questions can be answered by querying the csv file, then SQL agents would be the best choice. If your questions can be answered by vector search (answer and question have semantic relationship) rather than a logical query, then RAG can be the better approach. If you are not sure, test the approaches on a subset of the data and evaluate their performance. Although, my guess is that you would get the answer to your questions using logical queries, then agents would be my first suggestion.

@@airoundtable is agent more token consuming as I am using openai

@@harrysharma3435 Yes in general agentic systems are more token-consuming than RAG approaches. The number of agents collaborating together, the data size they have to process in every query, and the number of times they have to correct themselves to get the answer directly impact the cost of the agnetic systems.

I didnt't understand what rocama is. I might be able to help if you can give me more info

Thanks for the kind words. I am very happy to hear the content was useful for you. I might make another video to improve the SQL agent and add more features to it. But at the moment I am working on two other videos so that one will happen somewhere down the road. In the meantime, there are some articles that you can check to get an idea of how to make agents for data visualization. Here are some examples. I hope they help dev.to/ngonidzashe/chat-with-your-csv-visualize-your-data-with-langchain-and-streamlit-ej7 medium.com/@nageshmashette32/automate-data-analysis-with-langchain-3c0d97dec356 www.reddit.com/r/LangChain/comments/1d2qqqy/building_an_agent_for_data_visualization_plotly/

Thanks!

It is correct. You can find the answer to this question in row #15 of the titanic_small dataset: Pclass: 3 - Name: Mr. Anders Johan Andersson - Sex: male - Age: 39 - Siblings/Spouses Aboard: 1 - Parents/Children Aboard: 5 - Fare: 31.275

I don't have a video about combining them. But to combine them first there need to be a logical use case that needs a specific roadmap and mapping, since combining them together will not solve a general problem. But in case you have a usecase in mind you can definitely do it

You are right, llama-index is great and has a lot of tools. I will check it again soon. My last interaction with it was for a video in which I compared llama-index and Langchain for RAG. It was a while ago and I know they have evolved the framework alot.

Thank you. There aren’t enough llamas-index tutorials and I think you’ll explain them better than anyone! Learning so much from your videos.

@@sujit5013 I appreciate the kind words. I am working on two videos now. But I will keep llama-index inmind and check it out later for sure. Thanks again for the suggestion!

Thanks!

The code that I published on github does not have a memory. But yes it can have memory. You can use Langchain memory classes and integrate them into the chatbot. For start check the following link: python.langchain.com/v0.1/docs/modules/memory/agent_with_memory_in_db/

It depends on the use case. if there is a logic that a human can understand and decide then an LLM agent on top of both with that logic being passed to it can make the decision for your users. If the logic is only a matter of whether the sql agent fails to retrieve the answer, then you can add that logic in your code as well. But if there is no specific logic behind the selection, then that can't be integrated into the chatbot.

@@airoundtable I have seen some routing concepts recently. May be worth checking

I cannot say with certainty since I haven't seen the code behind AYD. and I am not sure what you meant by the "training chunks". But my guess is more toward the same approach that LangChain has been working on. LLMs are good at generating queries for these databases and by passing the schema of the database to the LLM it can understand how to retrieve the answer to the user's question. Next level would be to also add some necessary description of the database to the system prompt to help the LLM navigate easier. And finally, with the help of few shot learning you can do the final adjustment of the LLM on your database. If the database is huge, in a custom solution, I would break it down into smaller databases and utilize the divide and conquer strategy to achieve a good performance.

Thanks! I appreciate the kind words. I am glad that the video was helpful

Hi Darkmatter9583. I would be happy to help. You can go through the tutorials and ask your questions. I work on multiple projects at the moment but I will respond to questions whenever I can. In case you would like a head start in the field, send me a description of your background and what you want to accomplish. I will try to guide you in the right direction. You can send me a message on Linkedin.

Hello Darkmatter9583. Thanks! I answered your other message

Thanks! Does it have any free features? As far as I remember, Lamini is a commercial product. I am not planning to go through the commercial products unless they make a lot of sound or I get a sponsorship :)) But I will keep it in mind for a potential project down the line. Thanks for the suggestion!

The same agent that I designed and used in this video is dealing with multiple tables. It has access to around 11 tables and can easily get the answers. If you check the link below, you can see the diagram of the database that I am using in this video: docs.yugabyte.com/preview/sample-data/chinook/

Thanks. I am glad vyou liked the video. Yes the LLM is aware that it should not use its own knowledge and only provides information that exists in the database.

It is available: github.com/Farzad-R/LLM-Zero-to-Hundred/tree/master/RAG-GPT

Thanks!

I am adding the vector embeddings of my data to that vector index at 49:50. Then I use that vector index for RAG in this chatbot

Hi. if you are planning to use CSV for RAG please check out my video called "Chat with SQL and Tabular Databases using LLM Agents". The methods that I implemented in this video for llamaindex are only for PDF and text documents

I am working on it. Coming soon!

Thanks!

Update the agent strategy using this approach and check the results: python.langchain.com/v0.1/docs/use_cases/sql/large_db/

Thanks. I use draw.io

@@airoundtable Thanks. Just like we can ask numerical or statistical questions from thi chatbot, can we also ask questions (how and why) like why this number is high or low like explanation of an answer because LLM are good in generating text.

@@muhammadtalmeez3276 Not with this agent. This agent is designed to just retrieve information from databases. What you are looking for is another feature that needs to be added to the system for data analysis. That is a different problem and very challenging to pull out depending on the size and complexity of the database

What framework are you using? OpenAI from Azure or OpenAI directly?

Happy to hear it! Thanks

Thanks for the suggestion. The logic behind whether to use RAG or agents depends on the available database and the type of user's query. If the logic is known to us, it can be implemented by adding an LLM in the first step of the pipeline with the proper instruction to call two functions. One pointing to RAG and one pointing to the agent. If you are familiar with function callin, it would be a very easy addition. I also have multiple videos showing how function calling works. But I will use that technique again for sure down the line.

Thanks!

Yes, you can do that.

Thanks. For security and data privacy please have a look at AYD's blog here: www.askyourdatabase.com/blog/how-does-askyourdatabase-protect-your-data 2. If you want to connect to multiple databases individually, this is straightforward. However, if you aim to connect to multiple databases simultaneously, allowing the LLM agent to access and retrieve information from them at the same time, I didn't see that feature available. If your goal is to extract meaningful information using shared knowledge between these databases, this falls under the scope of a knowledge graph problem, which requires a tailored approach for each scenario. I personally haven't seen any open-source product like AYD. The only alternative approach that I am familiar with is to design the whole system from scratch.