Returning these multiple sources is great, but I’d like to see a workflow for refining this data into a single well structured long-form response. Some sort of hierarchical formatting of the key points in each data and then maybe recursively structuring that into a full document? I feel the chunking and context length of GPT4 tends to want to wrap each response with a “start”, “middle”, “end” : this is a difficult behaviour to break with prompting and makes it hard to use it to sequentially structure larger responses.
Could you cover RAG for data-heavy documents that contain a lot of tables. So not only storing the tables, but also pre-summarizing them to store the summary as embedding
I have huge financial documents.; about 200 pages long. And I have built a RAG based bot using LlamaIndex to answer questions from the documents. But it's not able to answer some questions; for example, who are the board members? What is the price of the PPA and its length? etc. Any suggestions on improving it? I was thinkking about hyDE or trying out different embedding models. But your suggestion would be appreciated.
hi james, let's say the content i'm retrieving from a vector db to pass as context to the llm (i am using gpt3.5) is large and i wanna generate questions based on this context but in my case gpt3.5 has max limit of 4096 tokens for which i had to iterate and change the context every time while calling this function "RetrievalQA.from_chain_type" which has retriever as an argument which takes the entire content it is retrieving as context but i want some of it to be passed as context every time or is there any other way ? i'm still a beginner
Thanks James, for the good tutorial! -I am using Faiss instead of pinecone with a local GPT embedding, having a relevance score from 0 to 1 for the findings. Would you prefer pinecone over faiss or doesn't it matter?
Hi, sir. Since you are working with sequential data, may I kindly suggest that you consider creating a video tutorial on implementing Transformers for time series data? This tutorial could cover topics such as forecasting, classification, or anomaly detection. It's not necessary to cover all of them; just one would be sufficient.
The challenge with RAG pipeline that i am facing is model hallucination. If my document contains two set of data one describing about certain functionality that uses a library and then a technical text on library itself, when question is asked on technical implementation of the libraries it falls back to the functionality and never explains about the implementation using the libraries which is in the doc which was embedded.
Any tips on solutions to test the performance of different chunking and embedding models for RAG alongside each other James. I recall seeing one but can't for the life of me find it!
If it was from my videos you might be looking for this ruclips.net/video/BD9TkvEsKwM/видео.html You can also use RAGAS - I haven’t used it myself yet but some of the team are on projects and seems to be helpful
Thanks for this video. Really helpful. I am trying to get results from multiple different documents that I have in my vector store. I want to reference the file name in the results that come back (source citation if you will). I see the metadata contains the source (which is the file name) and the page number so this should be possible. I am just not sure what technique to use to get the llm to cite these. I tried setting the prompt to do this but I did not get the result I wanted.
either in-memory or vector db. in-memory like faiss or llama index vector db like pinecone, pgvector, mongodb vector search, elasticsearch, milvus, qdrant.. etc
I am working in a project in Brazil and you are helping a lot! Thanks
Both of us! What type of project are you in? I'm just starting in this field
Returning these multiple sources is great, but I’d like to see a workflow for refining this data into a single well structured long-form response.
Some sort of hierarchical formatting of the key points in each data and then maybe recursively structuring that into a full document? I feel the chunking and context length of GPT4 tends to want to wrap each response with a “start”, “middle”, “end” : this is a difficult behaviour to break with prompting and makes it hard to use it to sequentially structure larger responses.
Could you cover RAG for data-heavy documents that contain a lot of tables. So not only storing the tables, but also pre-summarizing them to store the summary as embedding
convert the pdf to html
I've been doing this no-code with flowise for several months
yes afaik flowise is built on top of langchain.js - and I believe this is a relatively old retriever in LangChain :)
Can I build a RAG on flowise?
Hi, James, do you have any videos on how to evaluate/score performance of the retrieval documents with RAG?
it's an old one but yeah here ruclips.net/video/BD9TkvEsKwM/видео.html
I have huge financial documents.; about 200 pages long. And I have built a RAG based bot using LlamaIndex to answer questions from the documents. But it's not able to answer some questions; for example, who are the board members? What is the price of the PPA and its length? etc. Any suggestions on improving it?
I was thinkking about hyDE or trying out different embedding models. But your suggestion would be appreciated.
hi james, let's say the content i'm retrieving from a vector db to pass as context to the llm (i am using gpt3.5) is large and
i wanna generate questions based on this context but in my case gpt3.5 has max limit of 4096 tokens for which i had to iterate and
change the context every time while calling this function "RetrievalQA.from_chain_type" which has retriever as an argument which takes
the entire content it is retrieving as context but i want some of it to be passed as context every time or is there any other way ? i'm still a beginner
Thanks James, for the good tutorial! -I am using Faiss instead of pinecone with a local GPT embedding, having a relevance score from 0 to 1 for the findings. Would you prefer pinecone over faiss or doesn't it matter?
Hands on! Love it.
Hi, sir. Since you are working with sequential data, may I kindly suggest that you consider creating a video tutorial on implementing Transformers for time series data? This tutorial could cover topics such as forecasting, classification, or anomaly detection. It's not necessary to cover all of them; just one would be sufficient.
The challenge with RAG pipeline that i am facing is model hallucination. If my document contains two set of data one describing about certain functionality that uses a library and then a technical text on library itself, when question is asked on technical implementation of the libraries it falls back to the functionality and never explains about the implementation using the libraries which is in the doc which was embedded.
Do we have multi query api in langchain.js ?
Any tips on solutions to test the performance of different chunking and embedding models for RAG alongside each other James. I recall seeing one but can't for the life of me find it!
If it was from my videos you might be looking for this ruclips.net/video/BD9TkvEsKwM/видео.html
You can also use RAGAS - I haven’t used it myself yet but some of the team are on projects and seems to be helpful
Thanks for this video. Really helpful. I am trying to get results from multiple different documents that I have in my vector store. I want to reference the file name in the results that come back (source citation if you will). I see the metadata contains the source (which is the file name) and the page number so this should be possible. I am just not sure what technique to use to get the llm to cite these. I tried setting the prompt to do this but I did not get the result I wanted.
Thanks Jimmy. is it similar to RAG-Fusion ?
Brilliant Stuff!
thanks Shaheer :)
I need to do this in my node.JS GPT 3.5 Chatbot.
Thanks, James. When you said to re-rank the 50 docs, what method could be used for the re-ranking?
using reranking models (crossencoders), see here ruclips.net/video/Uh9bYiVrW_s/видео.html :)
I've heard this technique called RAG Fusion.
how to make bloke mistral 7b model faster with lang chain rag
The interviewer asked me about the different types of indexing what will be the answer?
either in-memory or vector db.
in-memory like faiss or llama index
vector db like pinecone, pgvector, mongodb vector search, elasticsearch, milvus, qdrant.. etc
indexing could also refer to type of index, like Flat, IVF, graph-based (like HNSW), etc - I'd probably ask the interviewer to clarify