LangChain - Advanced RAG Techniques for better Retrieval Performance
ฝัง
- เผยแพร่เมื่อ 4 ก.ค. 2024
- In this Video I will show you multiple techniques to improve RAG Applications. We will have a look at ParentDocumentRetrievers, MultiQueryRetrievers, Ensemble Retrievers, Document Compressors, Self-Querying and Time Weighted VectorStore Retrivers
Github: github.com/Coding-Crashkurse/...
Timestamps
0:00 Introduction
0:55 Chunksize Experiment
5:45 ParentDocumentRetriever
7:15 MultiQueryRetriever
10:18 Contextual Compression
15:35 Emsemble Retriever
17:29 Self-Querying Retriever
21:10 Time-weighted VectorStore Retriever
Many requested a follow-up video with an example - Two-Stage Retrieval with Cross-Encoders: th-cam.com/video/3w_D1L0F-uE/w-d-xo.html
Thank you so much for this tutorial! It is exactly the stuff I was looking for!
Great to hear that. Thanks for your comment
Great useful content, with clear explanation. 👍
Nice work! few new methods of Langchain I was not aware of :)
Already Love your content ❤
Would love to see you making Production Ready Chatbot Pt 2 along with deployment part. Thankyou for producing quality content for free.
Thank you! I currently work on a Udemy Course, which explains how to deploy a Production Grade Chatbot on Microsoft Azure. It´s not free, but only costs a few bucks 🙂. Will release it in January. But of course I will continue to do Videos on YT which are completely free.
@@codingcrashcourses8533 very cool. where is it :D
Absolutely fantastic
Thank you so much :)
Excellent information!! Thank you. Liked and Subscribed.
Nice! Will release a follow up video with a practical example on monday ;-)
Liked and subscribed. Spot on!
Thanks for great video of this topic.
can you also post some videos related to LoRA with any LLMs of your choice.
Thank you so much for making this video! You create valuable content. I just have one question. I'm currently utilizing the Azure Search Service, and I'm curious if it's feasible to integrate all the retrievers. I've attempted to use LangChain with it, but my options seem limited to searching with specific parameters and filters. Unfortunately, there's not a lot of information available on how to effectively use these retrievers in conjunction with the Azure Search Service.
I tried ACS before and also was not tooo happy with it. My biggest con is that ACS does not support the indexing API. I prefer Postgres/PgVector :)
Thank you so much this is really good stuff
Thanks for your comment :)
You're welcome I just bought your course actually 🙂@@codingcrashcourses8533
Thank you for the video:). In your opinion which method of retrieval will give me the most accurate output ( the cost is not as important in my case )? I work in pharma industry - tolerance to LMMs mistakes is very low.
I can not give you a blueprint for that. Just try it out and experiment. You know your data and there are so many different ways to improve performance. If cost does not matter the easiest way is use GPT-4 instead of GPT-3.5. Also try chain of thought prompting and then use one of the techniques I showed in the notebooks. There are so many ways to improve performance :)
Thanks for the video, what is x & y dim in the scatter plot (5.19)?
Tjw axis is of the plot
Fantastic video! :D
Quick question: Do you know how it's possible to create a local vector database that's queried via code, so the database doesn't get initialised each time the script is run?
Would really appreciate your help!
You just have the use the correct constructor for that Database class. Methods like from_documents are just helper functions to make that easier. Not sure if I understood your question correct though
Yeah, answered my question pretty much, thanks a lot! Do you know which function i can use to create a local database, that can also be passed to the SelfQueryRetriever.from_llm() constructor?@@codingcrashcourses8533
Thank you for the amazing tutorial! I was wondering, instead of using ChatOpenAi, how can I utilize a llama 2 model locally? Specifically, I couldn't find any implementation, for example, for contextual compression, where you pass compressor = LLMChainExtractor.from_llm(llm) with the ChatOpenAi (llm). How can I achieve this locally with llama 2? My use case involves private documents, so I'm looking for solutions using open-source LLMS.
Sorry, I only use the OpenAI models due to my old computer. Can´t really help you with that
Thank you, can you handle theproblem of retrieval when we ask question out of context of rag or greeting for exemple ?
You May check nvidia guardrails
I'm a beginner here and I've been using langchain from your videos. Is the advanced RAG instead of doing something like my code below where instead of using the search type as similarity, I'm using the types that you showed in the video yet everything else stays the same like using ConversationalRetrievalChain, prompt, memory etc...?
retriever=knowledge_base.as_retriever(search_type = "similarity_score_threshold", search_kwargs = {"score_threshold":0.8})
Also, which would you recommend to retrieve for large documents? I need to do RAG over 80 PDF documents and have been struggling with accuracy.
Lastly, in your OpenAi embeddings, why are you using chunk_size= 1 when by default, its chunk_size = 1000? Can you explain this part also please and thank you in advance
The advanced techniques also work with memory etc., but with the High Level chains I showed I may become a little bit difficult and "hacky".
In general I don´t set any scores, but just retrieve the best documents. I also don´t have an answer for setting a good threshold. In general I recommend using the get_documents method with the retriever interface for getting documents.
I set the chunk_size to 1 due to rate limit errors I often experienced. With higher chunk sizes it just makes too many requests at once it seems.
hi, in retrievalQa from langchain, we have a retriever that retrieves docs from a vector db and provides a context to the llm, let's say i'm using gpt3.5 whose max tokens is 4096... how do i handle huge context to be sent to it ? any suggestions will be appreciated
Gpt-3.5 Turbo allows 32 tokens I guess, gpt-4-turbo 128k. If you really need that large context window, my go-to apporach would be to use models with larger context windows at the end of 2023. There are also map-reduce methode to reduce the context, but these also do many requests before sending a final one.
Nice tutorial . May I know the theme used for visual studio code please
Material Theme dark :)
@@codingcrashcourses8533 link for the theme please as I see lot of material themes in market place extensions
Nice video. Can you please create a video on evaluation of RAG? I think a lot of people would be interested in this.
Thank you! That kind of video is currently not planned, since it´s actually quite expensive to evaluate RAG Output and designing that experiment is PROBABLY something not many people would watch on TH-cam. In addition to that I am not really an Expert on that topic. In my company our data scientists currently work on this^^
@@codingcrashcourses8533 what would be the right technique for evaluating a RAG? If you can share your thoughts in chat?
PDFInfoNotInstalledError: Unable to get page count. Is poppler installed and in PATH?
Though is it?
Wait what, I thought FAISS didnt support metadata filters ?
Weird that TimeWaited works with it no ?
I am not too familiar with each change, FAISS is also work in progress, maybe they added it in some version :)
@@codingcrashcourses8533 In any case, your video is amazing and you are greatly helping me for my internship project.
Many thanks, keep up the great work 💪👍