DocMesh
Semantic document intelligence · fully local · 2026
Upload a corpus and interrogate it: hybrid semantic + keyword search (FAISS dense vectors and BM25 fused with reciprocal rank fusion, then cross-encoder reranked), an automatically inferred graph linking documents by shared meaning, entities and topics, and grounded Q&A whose every citation is verified against what was actually retrieved. Every layer — chunking, embedding, fusion, reranking, graph inference, generation — is hand-built from primitives, no LangChain or LlamaIndex. It runs entirely on CPU with a quantised 1.5B local LLM, so no document ever leaves the machine — treating uploads and retrieved text as hostile input throughout.
No external LLM API · ~126 ms hybrid search · 3-signal doc graph
- FastAPI
- FAISS
- BM25
- Sentence Transformers
- spaCy
- llama.cpp
- React
