scikitplot.mcp

Composable documentation retrieval for Model Context Protocol servers.

The default import surface is independent of the MCP SDK, pydantic, and the optional corpus/vector dependencies. Importing scikitplot.mcp pulls only the SDK-free retrieval core (contracts, fusion, the in-memory demo retriever) and the pydantic-free capability/runtime-status helpers, so it works on the Legacy Retrieval tier (Python 3.8+) with a base install.

The server layer (SearchService, create_server, and the pydantic output models SearchDocsOutput / CitationOutput) is imported lazily on first access. It requires the [mcp] extra (pydantic always; mcp>=2.0.0,<3 on Python >= 3.10). No models, indexes, files, or network connections are opened at import time.

See also

• modelcontextprotocol/python-sdk

• semantica-agi/semantica

User guide. See the Model Context Protocol (MCP) section for further details.

MCP: Model Context Protocol

User guide. See the Model Context Protocol (MCP) section for further details.

Class inheritance

DocsRetriever	Structural contract every retrieval backend must satisfy.
RetrievedChunk	One retrieved passage with the metadata needed to cite it.
InMemoryBm25Retriever	A compact BM25 implementation suitable for demos and small corpora.
Bm25Retriever	Lexical retriever (FTS/BM25) leg backed by a full-text search seam.
CorpusAnnoyRetriever	Docs dense retriever backed by an embedder + a vector index + a document lookup.
HybridRetriever	Fuse several retrievers into one via Reciprocal Rank Fusion.
build_search_docs_result	Format retrieval results as an MCP tools/call response with citations.
builtin_demo_retriever	Return a tiny corpus that explains the sample's own mechanism.
create_server	Create an official MCP Python SDK v2 MCPServer instance.
reciprocal_rank_fusion	Fuse weighted ranked lists into a single key -> score map.