Agent-native retrieval

Indexes retrofitted from lazy human search behaviour are tuned for the wrong workload. Agent queries are goal-driven, highly specific, quoted and filtered — and there would be far more of them if each one were cheaper. The cost of a query is currently what caps how much an agent searches, which is a strange place for an economic ceiling to sit: the model gives a confident-but-average answer because it cannot afford to go and check.

The sharper failure is not that the right passage is missing but that nothing separates it from a near-miss. On a specialist corpus a leading commercial embedding model ranked the correct passage far down the list with near-zero score separation from a plausible wrong one. The reported fix is not more domain text or a bigger corpus but task-specific, ontologically structured contrastive training — which reframes the whole thing as an alignment failure rather than a recall shortfall. What is being sold alongside it is trust, and trust loss is asymmetric: one confident near-miss costs more than many correct answers earn.

A third strand is blunt operational reality: platforms and APIs offer no first-class entrance for agents and break at agent request rates. Practitioners respond by building their own door — a local index, purpose-built full-text search over the corpus the agent must reason across — instead of hammering a human-paced API.

Assay retrieves by path, filename and grep today — briefs, skills, findings and memory are addressed structurally rather than by similarity. That is an accident of design age rather than a considered rejection, and it will change the moment any similarity-based feature is proposed. Two candidates are obvious: deduplication at the front door ("is this issue the same as that one?") and a citation matcher for an out-of-band learning pass. Both are exactly the near-miss-versus-hit problem.

It is procedural: before trusting any similarity score over Assay's own corpus, build a discrimination benchmark from Assay's own ontology — desk roles, gate types, the stream taxonomy — and measure whether an off-the-shelf embedding can tell a near-miss from a hit on it. Off-the-shelf performance on a specialist vocabulary should be assumed near-zero until measured, and the failure mode is the dangerous kind: a confident citation of the wrong finding is worse than no citation, and it degrades exactly the evidence chain the methodology exists to protect. This sits alongside the standing rule that invariants stay closed-form checks rather than model judgment.

The desks' corpus is small and local, so search is already cheap and there is no economic ceiling on how much a worker greps — worth noting only as a reason this pattern may not generalize to adopters with large private corpora. The local-index move does bite, though: the desks re-fetch tracker and PR state from the forge API every loop, and a local index over the issue and PR corpus is the named alternative for when API pace or rate limits become the binding constraint. Rate-limit blindness is already a recorded desk failure mode.

• Whether a second domain reports the same near-zero within-domain discrimination from commercial embedding APIs — that would move this from a specialist result toward a general finding.
• Whether anyone publishes a cheap recipe for the ontology-structured contrastive training step, as opposed to a vendor result; that is the difference between "build a benchmark first" and "build a model first".
• Whether agent-shaped query interfaces — structured web query languages, dated evidence extraction — reach a usable implementation rather than a pitch.