Advance RAG control flow with Mistral and LangChain: Corrective RAG, Self-RAG, Adaptive RAG

Рет қаралды 15,390

Mistral AI

Күн бұрын

github.com/mis...

Пікірлер: 26

@nikitakuznetsov4592 7 ай бұрын

Guys, this is crazy good! Please don't stop your demos and explaining of concepts. If you read this - can you explain a lil bit more the concepts of action tools (usage, own implementations and so on). Thx in advance!

@bqmac43 7 ай бұрын

Tools are functions that the agent can call. To decide which tool to use, an agent can send the available tools to the LLM and say "which one should I use?" Once a tool has been selected, the LLM can then provide arguments to pass into the tool's function. The agent takes the information from the LLM to call the tool, and then goes back to the LLM to say "What tool should I use now?" In the example that is shown here, the agent is given specific routes to take. This simplifies each step because the agent is focused on a specific outcome at each step. So at each step, the available tools are scoped down to the task at hand. An alternative to this flow are ReAct agents. ReAct agents are given a set of tools and a task and can reason for themselves how to accomplish the task given the tools they have. Each type of flow has it's place (as Lance points out nicely with his pros and cons). Personally, I start with ReAct agents because they're easier to set up and if I feel myself getting frustrated by the steps it takes, then I move to a more deterministic flow (i.e. LangGraph, what Lance does in the video). That's a long explanation and hopefully it makes sense. You can read more on how to implement them with Langchain here. python.langchain.com/docs/modules/tools/custom_tools/

@choiswimmer 6 ай бұрын

The langchain channel has more

@nicolaspellerin2207 6 ай бұрын

Thanks for this ! Learned a ton of good stuff, very well explained, will definitely be playing with your notebooks 😊 You’re fantastic for sharing such high quality work

@NarendraChennamsetty 7 ай бұрын

This is an amazing tutorial. so much valuable information packed in 30 min. Subscribed, Thank you!

@bqmac43 7 ай бұрын

Lance's videos always have great insights. I'd recommend checking more of his videos out if you liked this one.

@RajaSekharaReddyKaluri 6 ай бұрын

Thank you Sophia and Lance!

@hxxzxtf 5 ай бұрын

🎯 Key points for quick navigation: 00:00 *- Advance RAG control flow with Mistral and LangChain* 00:12 *- Combining small steps into comprehensive control flow for large language model applications* 00:25 *- Flow engineering uses a flow diagram to check response intent and construct answer iteratively* 01:06 *- Corrective RAG uses retrieval evaluator to assess document quality and trigger web search for additional information* 02:14 *- Hallucination note checks answer support by document, and answer question node checks generated answer relevance* 10:31 *- Bind MRAW to schema* 10:43 *- Convert JSON output* 10:59 *- Mock retrieval example* 11:12 *- Grading documents relevance* 11:25 *- Confirm binary score* 11:39 *- Define RAG chain* 12:05 *- Graph State explained* 21:11 *Adversarial Tax Routing* 21:52 *Hallucination Grader Defined* 22:18 *Router Conditional Edge* 22:47 *Web Search Fallback* 24:03 *Control Flow Implemented* Made with HARPA AI

@8eck 6 ай бұрын

Glad to see you in Mistral AI! 🥰

@AlbertJinkuGu 6 ай бұрын

Awesome job! Thank you for sharing! What's the best way to do the RAG based on the relational database? We need to understand the question, go to the correct table of a database and find the most relevant records. Looks like we should support both keyword search and sematic search. For the keyword search, we need to extract the parameters, like the keyword, date of that question, the person who generated that record, etc.

@TheInternet81 3 ай бұрын

but better if you can store several version of perspective and has calculation the benefit do a perspective than other perspective. because...in academic world there is several perspective to solve a problem. what you build here is only enhancement a perspective. event that we all here should appreciate this is a BIG STEP forward of improvement in field AI Knowledge. cheers...

@kuldeepsinhjadeja3668 5 ай бұрын

In the last part, when the flow went twice to the web search tool, it basically searched on the same query, then how did it produce valid result 2nd time and not first time. How to ensure that it does not get stuck in the loop, because basically it does the same thing again and again without changing anything hoping to get correct result.

@davidvukotic 3 ай бұрын

The best 💯

@awakenwithoutcoffee 4 ай бұрын

thank you for the wonderful insights in the latest RAG developments. Can someone explain in simple terms the benefit of implementing "LangGraph" ? from what I understand it allows for more accurate LLM executions by limiting the "routes" the output of a certain LLM flows trough, improving it's reliability in execution. But why can't we empower LangChain "Agents" with the same functionality ? wouldn't the ideal agent have LangGraph capibilities built in ?

@awakenwithoutcoffee 4 ай бұрын

after diving deep for 3 weeks I know the answer to my own question (lol): - LangGraph is an extension to LangChain that allows for managing "state" e.g.: controlling what goes in and out your LLM. By restricting what the LLM is able to do we increase it's reliability allowing us to build better, safer & more complex LLM systems. Instead of making 1 massive prompt technique we utilize the techniques of traditional System Design to map out our software in a more traditional sense.

@Taskade 6 ай бұрын

Can't wait to incorporate Mistral into Taskade in our next Multi-Agent update :)

@camaycama7479 6 ай бұрын

Keep us posted!

@choiswimmer 6 ай бұрын

Is this an ad? It's an ad

@aipt32 5 ай бұрын

What happens if the graph gets stuck in a loop? (Web Search > not use full > Web Search > not usefull > ...) Do i have to add a "tries" counter to my state and end after x tries to prevent an infinte loop?