[WIP] Build advanced RAG architectures with Mistral-7b using LlamaIndex

https://huggingface.co/learn/cookbook/en/rag_evaluation https://python.langchain.com/v0.1/docs/integrations/chat/google_vertex_ai_palm/ https://cloud.google.com/vertex-ai/docs/tutorials/jupyter-notebooks#vertex-ai-workbench

start with:

1. what is rag?
   • how is it different than fie-tuning?
   • ways of evaluating rag
     • humans
     • static scores
       • precision
       • recall
       • bleu, rouge
     • use another llm
       • since broad capabilities
       • varied human preferences

2. lots of things to tweak in the RAG pipeline - but there should be a proper way to evaluate its impact

3. dataset used is huggingface documentation
   • text + source
4. Build RAG
   • split documents using recursive split
   • embed documents
     • model used: thenlper/gte-small
   • retrieve relevant chunks - like a search engine
     • FAISS index stores embeddings for quick retrievals of similar chunks
   • use retrieved contexts + query to formulate answer
     • rerank retrieved context
       • retrieve 30 docs
       • rerank and output best 7
       • model used: colbert-ir/colbertv2.0
       • reevaluates and reorders the documents retrieved by the initial search based on their relevance to the query
     • answer using llm
       • LLM B (zephyr-7b-beta)
       • a fine-tuned version of (mistralai/Mistral-7B-v0.1) trained to act as helpful assistants.

basic rag sentence-window rag auto retrieval rag