Model Context Protocol Retrieval#
scikitplot.mcp is a documentation-retrieval and MCP adapter layer. The base import remains independent of the MCP SDK and pydantic; retrieval can be built and tested first, then exposed through the official MCP server layer when that optional tier is installed.
Important
MCP owns wire/adapter semantics, not evidence truth. Corpus owns ingestion and retrieval data semantics, Annoy owns vector mechanics, and the official MCP SDK owns the protocol implementation. Keep those boundaries explicit.
Use the dependency-light retrieval core before adding an MCP server.
Separate retrieval semantics, MCP wire/server semantics, Corpus, and vector index ownership.
BM25/demo retrieval, hybrid retrieval, reciprocal-rank fusion, statuses, and deterministic truncation.
Compose Corpus evidence and vector search without moving ownership into MCP.
Inspect advertised/effective server capabilities and runtime status.
Lazy SDK/pydantic server layer, stdio/network transport, validation, limits, and remote-bind grants.
Local stdio/HTTP, Docker, remote gateways, health checks, and runtime controls.
Core contracts, retrievers, outcomes, server exports, and validation helpers.
Optional SDKs, capability mismatches, degraded retrieval legs, and server startup.
Use the dependency-light retrieval core before adding an MCP server.
Which page should I use?#
Start with Getting started. Then choose the page that owns the decision you are making: architecture and lifecycle, configuration, persistence, performance, public API, or limits. The guide intentionally separates those concerns so examples do not accidentally become policy.
Examples#
See Mcp for executable examples that accompany this guide.