What it does
Tiburcio indexes a codebase — source files, architecture docs, team standards — into a Qdrant vector database. Developers can then ask questions and get answers grounded in their actual code, not generic suggestions.
It ships in two forms: a web chat UI built with Vue 3 and Hono, and an MCP server so Claude Code can query the same knowledge base while you’re coding.
How it works
Documents go through chunking and embedding, then get stored in Qdrant. At query time it uses hybrid search — BM25 for keyword matching plus dense vectors for semantic similarity, fused together with RRF. Indexing runs on a schedule through BullMQ job queues backed by Redis.
The whole thing is containerized with Docker. PostgreSQL (via Drizzle ORM) handles metadata and user state. There are 162 automated tests covering the core flows.
Why I built it
I wanted a way to onboard faster on large codebases at work. The existing tools either hallucinated too much or didn’t understand project-specific conventions. Started as a learning project to dig into RAG and vector search, then grew into something actually useful.
It’s MIT licensed, currently at v1.2.1, and being adopted at Prozis for internal use.