janssen/claude-memory/project_mcp_soubory.md

name, description, metadata

name	description	metadata
project-mcp-soubory	MCP server "soubory" - hybridni dotaz nad PG fulltextem a Mongo metadaty pro Dropbox souboru obou studii	node_type type originSessionId memory project 49aa480f-6667-4832-b091-08f333c27872

MCP server soubory v mcp_soubory.py, registrovan v U:\janssen\.mcp.json jako soubory.

Architektura: Postgres MongoSoubory.documents (fulltext tsvector + body) + MongoDB soubory.{42847922MDD3003,77242113UCO3001} (metadata + content.*). Source skripty: scan_files_v1.0.py, enrich_files_v1.0.py, enrich_fulltext_v1.2.py.

Tools:

• ping - health check + counts per studie
• list_studies - prehled korpusu, ext breakdown, fulltext coverage
• search(query, study?, ext?, since?, folder?, limit, with_metadata) - HLAVNI fulltext (websearch_to_tsquery), ranked + ts_headline snippet <<...>> + Mongo content.* enrichment
• read_document(path|mongo_id, offset, length, around_match) - cele body, slice nebo okno kolem matche
• get_metadata(path) - raw Mongo doc bez body
• recent_files(study?, days, ext?, limit) - co se zmenilo (mtime)
• find_duplicates(study?, min_size_kb, limit) - sha256 groups, wasted_mb_estimate
• by_author(name, study?, ext?) - regex na content.author / last_modified_by
• browse_folder(folder, study?, ext?) - regex na parent_folders

Pro EML/MSG pouzij search(..., ext=["eml","msg"]) + by_author(jmeno, ext=["eml","msg"]) - samostatny email tool je redundantni a navic by se pletl s project_graph_email_import.

Aliasy studii: MDD3003 -> 42847922MDD3003, UCO3001 -> 77242113UCO3001. None = obe.

Why: project_claude_learning - chce se ptat v chatu ("najdi mi dokument o X", "co psal sponsor Y"), nikoli volat CLI.

How to apply: Po restartu Claude Code budou nastroje dostupne jako mcp__soubory__*. Pred volanim search radeji nejdriv list_studies aby clovek vedel co je v korpusu. Pri velkem body pouzij around_match misto stahovani celeho dokumentu.