fix(loki): find the contact that IS there — the refusal was a false negative

[catomean]

Browser-tested the shipped harness (#212, #214) against the exact question that started all of this:

"Who is Ilya Druzhnikov in my contacts, and what is his affiliation? Also do research on Elena Weber — linkedin etc."

The good news: no fabrication. No UZH. No invented biography. Both halves refused cleanly:

Regarding Ilya Druzhnikov: Not in your data.
Regarding Elena Weber: I cannot browse the web or conduct
external research. Not in your data.

The bad news: that refusal was wrong. Ilya Druzhnikov is one of 1285 rows in the people table.

Hardening against fabrication moved the failure to the opposite end — and this end fails quietly, because a confident refusal looks like diligence. A grounded assistant that cannot find real records is not safer than one that invents them; it is differently useless.

Two bugs, both found only by driving the real UI

1. peopleFacts passed the whole message as a name filter. searchPeople filters with name ILIKE '%q%', so a sentence can never match a name column. Every lookup silently fell through to "12 most recent contacts", and the named person was simply absent from the context — after which "Not in your data" was the correct answer to a context that was wrong.

Now candidate names are extracted and searched individually. That extraction had its own bug, caught by a test before shipping: matching only capitalised tokens makes the lowercase words between two names invisible, so "Ilya Grün the same person as Jean-Luc" merged into one nonsense run. Fixed by tokenising every word and breaking the run on anything that isn't a capitalised non-stopword.

2. A groq 429 abandoned the tool loop. The box log:

[loki] tool loop failed: groq 429: Rate limit reached for
model `llama-3.3-70b-versatile` ... service tier `on_demand`

It fell back to the gateway — a path with weaker retrieval. A rate limit is a capacity problem, not a capability one, so a capacity blip was silently degrading answer quality. It now steps down to llama-3.1-8b-instant, already verified to drive the loop via the text protocol (npm run probe:models).

Verification

scripts/test/agent-name-lookup.ts pins both, including the exact production message, plus first-name-only lookups, stopword rejection, accented (Ilya Grün) and hyphenated (Jean-Luc) names, and the 429 step-down. 67/67 unit files pass, tsc clean, 0 lint errors.

🤖 Generated with Claude Code