Corpus and Annoy#

Use Corpus when evidence originates from real files, URLs, archives or media. MCP should receive retrieval-ready evidence rather than become another ingestion framework.

CorpusIndexRetriever composes a Corpus-style index with optional embeddings and supports lexical/vector/hybrid matching. CorpusAnnoyRetriever adapts an embedder plus vector index and document lookup function.

Annoy results are approximate similarity candidates. Preserve source metadata and citations through the adapter so a client can trace a returned chunk back to the evidence that produced it.