Watch MultiVault actually work

▸ Setup, one time

Vault points at ~/Work/company-ops (41 files: 26 project docs, 12 email exports, 1 team.ics). First index: 2,847 chunks in 6.1s. Background watcher on. Nothing leaves the machine.

▸ The query, typed in Claude Desktop

"Why did we kill the Q2 pricing model? Sarah objected somewhere in a thread."

Claude calls get_context automatically. MultiVault runs BM25 over the exact terms, runs the local semantic pass in parallel, and fuses the two lists with Reciprocal Rank Fusion. Five chunks come back:

1. email-exports/sales-thread-mar2024.txt — BM25 rank 1, semantic rank 4
"Re: Q2 pricing v3 — Sarah Okonkwo, Mar 14 09:41: I can't sign off on this. The usage tier floor is the problem, not the headline number."

2. docs/pricing-model-q2-v3-review.md — BM25 rank 2, semantic rank 2
Section "Open objections," bullet 3, unresolved.

3. notes/leadership-sync-0318.md — BM25 rank 47 (near-miss on exact terms), semantic rank 1, fused to position 3
"Moving three of our top five accounts onto metered billing strands them mid-contract. We'd be renegotiating in Q3 with no leverage. Agreed to shelve."

That third one is the actual reason, and it never uses the words "pricing," "Q2," or "reject." Keyword search alone ranks it 47th. Semantic mode is what surfaces it.

4. team.ics — "Pricing decision — final" Mar 18, 14:00, 5 attendees. Timestamps the shelving.

5. email-exports/finance-inbox-q1.txt — Sarah's follow-up, Mar 19, one line of confirmation.

vault_status: 41 files, index current as of 14 seconds ago, 0 stale.
MultiWitness (if