Open source2026

Sextant — MCP-Native RAG

An MCP-native retrieval-augmented generation system with cost-aware experiments, table-aware chunking, and query decomposition. Open source (MIT).

PythonMCPQdrantGeminiRAG

Code

Sextant is a ground-up, MCP-native RAG system built to push retrieval quality while staying under a strict cost budget. It is open source under the MIT license at github.com/hulagerushikesh/sextant.

The idea

Instead of wrapping retrieval in a bespoke API, Sextant exposes it as Model Context Protocol tools — so any MCP client (an IDE assistant, an agent, a chat app) can query the knowledge base natively, with no glue code.

What went into it

Cost-controlled experiments — every retrieval idea was run as a budgeted experiment measured in real dollars; table-aware chunking and a query-decomposition prompt earned their place and shipped, while others (HyDE, prompt-caching) were measured and deliberately rejected.
Correctness under change — per-document caps bound how much any one source can dominate results, and re-ingest-as-replacement semantics (plus a forget command) prevent stale chunks from lingering in the index after an update.
Modern UI — the front end was ported to React 19 + TypeScript + Tailwind v4 + shadcn.

Stack

Python, the MCP server SDK, Qdrant for vectors, and Gemini for generation — all under a configurable daily spend cap.