07 · Origin RAG system
ai-rag
RAG chat-with-your-documents: PDF ingestion, Mistral embeddings, ChromaDB retrieval with citations, Streamlit UI, LLM-as-judge eval.
ai-rag-project is a local RAG loop: PDFs are sentence-aware chunked, embedded with Mistral, stored in ChromaDB, and retrieved with hybrid search (vector + BM25, fused with RRF). Answers stream with inline citations back to source chunks.
A Streamlit UI handles upload and chat. An evaluation script scores Q/A pairs, including unanswerable trick questions. Later work in the same repo moved eval from 4/6 to 6/6 by fixing chunking, retrieval, and prompt mistakes the suite actually caught.
CareerAgent and RetrievalLab sit downstream of this project. The point of keeping it public is the lineage: citations and evals were requirements from the first system, not a retrofit.
In the repo
- Sentence-aware chunking with overlap
- Hybrid retrieval fused with RRF
- Streaming chat with conversation memory
- Inline citations to source chunks
- LLM-as-judge eval with unanswerable traps