personal-notes-assistant CLAUDE.md

personal-notes-assistant

A FastMCP server that indexes an Obsidian vault into Milvus and answers questions over it via RAG (Ollama or OpenAI as the LLM backend).

Architecture

• main.py — module-level FastMCP("obsidian-rag") instance. Env vars are read and validated at import time (fails fast if OBSIDIAN_VAULT_PATH is unset). The lifespan async context manager builds one ObsidianRAGServer, runs its initialize() (Milvus connect, embedding model load, file watcher start, initial full ingest), and yields it as {"rag": rag_server} — tools reach it via ctx.lifespan_context["rag"].
• mcp_handlers.py — all MCP tools, @mcp.tool-decorated functions with type hints (FastMCP infers the JSON schema; don't hand-write one). _rag(ctx) is the one place that pulls the server out of the lifespan context.
• rag_server.py — all real logic: markdown/frontmatter parsing, chunking, Milvus CRUD, the watchdog file watcher, and the RAG query against the configured LLM client. This is where behavior changes belong; the tool layer should stay a thin pass-through.

Adding a tool

Add a function to register_tools() in mcp_handlers.py:

@mcp.tool
async def my_tool(ctx: Context, arg: str, limit: int = 5) -> str:
    """One-line description — shown to the LLM as the tool description."""
    return await _rag(ctx).some_method(arg, limit)

No manual inputSchema, no dispatch if/elif — FastMCP derives both name and schema from the function signature and docstring. Keep the tool body a one-line delegation into ObsidianRAGServer; put actual logic in rag_server.py.

Milvus notes (non-obvious, bit people twice already)

• collection.insert() must be followed by collection.flush(). Without it, a note just ingested/updated is invisible to search_notes and get_note_content — this was a real bug, verified against a live instance.
• Any user-controlled string (file path, filename) interpolated into a Milvus boolean expr filter must go through _escape_milvus_string() first — a bare " in a filename otherwise breaks or manipulates the query.
• The file watcher's on_deleted handler matters: without it, deleting a note from the vault leaves its chunks in Milvus forever. on_modified, on_created, and on_deleted all point at the same _handle — don't reintroduce three copies of the same body.
• Using the ORM-style pymilvus API (connections.connect, Collection, utility) — it's deprecated upstream in favor of MilvusClient, but that's a separate migration, not folded into unrelated changes.