Qdrant plus FastMCP is a build-your-own option, not a ready repository search service. Qdrant supports hybrid and multivector retrieval; the application still needs Git ingestion, Markdown processing, source reading, and a restricted public MCP interface.

A custom retrieval stack when control justifies the work

Qdrant's hybrid retrieval tutorial demonstrates dense and sparse retrieval, rank fusion, and an alternative ColBERT late-interaction reranking stage. FastMCP's HTTP deployment documentation covers serving tools remotely and deployment concerns.

My proposal is to combine these components only when corpus scale, concurrency requirements, custom ranking, or detailed filtering make a dedicated service worthwhile. This is not an assertion that they outperform the packaged candidates on the Operator's corpus.

The missing application work is substantial: repository synchronization, structural Markdown chunks, stable source identifiers, deletions, document hydration, query budgets, and a restricted search/read interface. A vector database and a protocol server do not supply that whole application merely by being installed together. No combined deployment or benchmark was performed in this exchange.