Improved-RAG-Architecture
Improved RAG Architecture using semantic chunker, query input rewriter, and prompt engineering
Stars
1
Committers
1
Improved-RAG-Architecture
Using LangChain as the framework Improved RAG Architecture from my previous project, such as
- Rather than running the model localy. I'd rather using API from together.ai so I don't destroy my laptop in doing so. (they also provide better model so the outcome of the prompt is wayyyy better). I could also tune the parameter of the LLM for better outcome or maybe model selection for my specific use case.
- Fixing the chunking problem by using LLM (Semantic Chunker) as the chunker rather than manual chunking.
- Rewrite the query using LLM before turning it into vector for retrieval purpose for better retrieval.
- Using both Semantic Search (Context) and Lexical Search (Keyword) for the vector DB, which is FAISS (Facebook AI Similarity Search).
- Reranking+autocut Algorithm after the retrieval for better output.
RAGAS could also be implemented in this project, a RAG architecture benchmarking There are other method of RAG I really interested in like GraphRAG.
P.S.
- My semantic chunker run verrry slow, i advise you just use normal text splitter with fixed size and overlap for faster performance
- If you don't want to host the db in local, you can use services like Pinecone or MongoDB Atlas and create a cluster
- I advise you to use other model for the inference and rewriter, find the best one for your usecase
Skibidi Sigma Rizz +1000 Aura