03 / 06
.
AI document intelligence platform
- Category
- RAG · LLM Evaluation
- Role
- Solo, end to end
- Link
- View on GitHub

By the numbers
- 0.61 → 0.89
- faithfulness after reranking
- BM25 + vec
- hybrid retrieval
- CI
- gated evaluation
Overview
A RAG system I built end to end. Retrieval is hybrid, BM25 plus vector search, then reranked with a Cohere cross-encoder. Quality is measured, not guessed: a Ragas evaluation pipeline runs in GitHub Actions and gates every change. The app is multi-user, with Google and X sign-in through NextAuth.
Highlights
- 01Hybrid BM25 + vector retrieval with Cohere cross-encoder reranking.
- 02Every answer cites the exact page it came from.
- 03Ragas evaluation pipeline, CI-gated in GitHub Actions.
- 04Multi-user architecture with NextAuth (Google & X OAuth).
What’s next
Next up: fully local inference with Ollama, with zero data leaving the machine.