Add MCP server

[MartinDevi]

Summary

• Adds a FastMCP server exposing the document index over the Model Context Protocol
• Tools: search, ingest, open_source, navigate_source, read_source, grep_source
• Supports both stdio (default) and HTTP transport modes
• Wires up MCP config via .mcp.json.jinja and documents usage in the README

Test plan

• Run make mcp and verify the server starts in stdio mode
• Run make mcp-http and verify the server starts on localhost:8000
• Ingest a document via the ingest tool and confirm chunks are indexed
• Run search and confirm results are returned
• Use open_source / navigate_source / read_source / grep_source on a returned hit

🤖 Generated with Claude Code

[haralambievk]

love the direction, but as this is a public repo let's coordinate with product when and how to put this out

[MartinDevi]

MCP Search Tools — Test Results

Tested all five MCP tools against two ingested documents:

• ACTIVISIONBLIZZARD_2015_10K.pdf — 182-page SEC 10-K filing (244 chunks)
• A 4-page French maritime safety guide from mer.gouv.fr (4 chunks), later deleted

search — hybrid BM25 + vector retrieval

Ran queries in both English and French. Results were correctly ranked with meaningful scores and pointed to the right sections of the right documents:

• "executive compensation CEO salary" → top hit: Part III / Item 11 Executive Compensation section (score 6.2), followed by the Stock-Based Compensation note (score 4.6)
• "équipements de sécurité balisage lumineux" → top 3 hits all from the French maritime doc (scores 14.2, 13.8, 13.4), zero bleed from the 10-K

Cross-document isolation worked correctly — semantically distant queries did not surface results from the wrong document.

grep_source — lexical phrase/term search within a document

• "Call of Duty" in the 10-K → 3 relevant chunks found across pages 5, 6, and 8
• "gilet de sauvetage" in the maritime doc → correct chunk found on page 3

navigate_source — sequential chunk navigation

Navigated forward 2 chunks from the Executive Compensation section (page 89): correctly returned Part IV (page 90) then the Signatures page (page 91) in reading order.

open_source — context window expansion

Expanded window=2 around the Executive Compensation chunk: returned the 2 preceding chunks (Items 9A/9B, internal controls — pages 87–88) and 2 following chunks (Part IV, Signatures — pages 90–91). Order and content correct.

read_source — offset-range reads

• Full document read of the 4-page maritime doc (top_k=4): returned all 4 chunks in order
• Range read of the 10-K with start_offset=334969, end_offset=340000: returned the 2 chunks that fall within that range (pages 89–90), nothing outside it

delete + verification

Deleted the maritime doc by source_id. Verified:

• grep_source on the deleted source → Source not found error
• search with a French maritime query → zero results from the deleted doc; only low-score 10-K results remain

All tools working correctly.

[alexandretamborrino]

I guess we also want to add exclude_ids as a parameter of the Search tool