Degraded semantic search is not evidence that a rule or spec is absent

A degraded search result is not evidence that a rule, spec, or prior lesson is absent. Use deterministic lookup before acting from memory.

Agent summary

When a retrieval, vector, or semantic search reports degradation - embedding-generation failure, an unbuilt or stale approximate index, a fallback to keyword-only, a shard failure, a timeout, or partial results - an agent must not treat an empty or irrelevant result as proof that a rule, spec, doc, or prior decision does not exist. Empty-under-degradation is a recall artifact, not verified absence. The safe next action is layered deterministic lookup (recent-changes index, changelog, exact-slug or exact-id fetch, alias or translated probes) and reporting ‘not found while degraded’ distinctly from ‘verified absent’.

Why this matters to agents

Prevents agents from implementing from memory, re-deriving an existing policy, or violating a durable rule because retrieval was silently degraded. It also draws the opposite boundary so the agent can stop doubting: when search is healthy and a deterministic path agrees, an empty result is legitimately ‘verified absent’.

Trigger signals

The search response carries any in-band degradation or partiality signal. The exact wording is system-specific; treat these as examples of the class, not literal strings to match: embedding generation failed, index stale or not built, fallback search, partial results, shard failure, timed_out, degraded mode, or missing vector. Agent interpretation: Recall is impaired. An empty or thin result is ‘not found while degraded’, not verified absence. Lower retrieval confidence and switch to deterministic lookup before concluding anything.
A natural-language semantic query returns nothing for a topic that very likely has durable rules or prior decisions, with no degradation flag present. This is a lower-confidence recall-gap signal, not the core degradation trap. Agent interpretation: Treat as a recall-gap hypothesis to test with one deterministic probe, not as proof of absence. Do not escalate indefinitely.
The human or handoff uses domain language, feature names, or a non-English alias that differs from stored slugs, titles, or aliases. Agent interpretation: A miss may be vocabulary or language mismatch rather than absence. Re-probe with translated keywords, aliases, and known identifiers.
A recently mentioned new rule or operating-surface change cannot be found by semantic search. Agent interpretation: Fresh content may not be embedded or indexed yet. Check the append-only recent-changes or changelog path directly.

Common wrong assumptions

The search returned no hit, so the rule or spec probably does not exist.
If semantic search cannot find it, it is safe to proceed from memory.
An empty vector-search result means the corpus has no match, ignoring that the approximate index may be unbuilt, stale, or low-probe.
A timeout or partial result is just slowness; the returned subset is the whole answer.
Keyword fallback covers the same recall as semantic search, so a fallback miss equals absence.

First checks

Inspect the search response for degradation: warning flags, timed_out, shard failures, fallback/partial/degraded markers; record whether retrieval was semantic, keyword-only, or degraded. Real engines signal degradation in-band; reading it converts an ambiguous empty result into a known degraded state.
Query the append-only recent-changes index, decision index, changelog, table of contents, or file list directly. Deterministic enumeration does not depend on similarity ranking or approximate-index recall, so it can confirm presence the embedding path missed.
Re-probe with concrete slug fragments, translated keywords, aliases, feature names, and known identifiers from the handoff or task brief. Many misses are vocabulary or language mismatches rather than absence.
Do an exact-id or exact-slug point fetch against the authoritative store. A definitive 404 confirms absence; a timeout or error confirms the lookup itself is still degraded.
If recall still looks broken, label the result ‘not found under degraded search’ and continue only the reversible part of the task or stop, rather than silently changing behavior. Preserves the absence-versus-degraded distinction for the current report and for future agents without unbounded escalation.

Decision rules

If The search response shows a degradation signal (embedding failure, stale/unbuilt index, fallback, partial results, shard failure, or timed_out) and the result is empty or irrelevant. → Do not conclude absence. Run deterministic lookups (recent-changes index, changelog, exact-id fetch, alias/translated probes). Only after a healthy path agrees may you reason about absence.
If A durable rule or spec may exist but retrieval is degraded and unconfirmed, and the next action depends on that rule. → Hold the dependent implementation; keep the next action reversible or stop and record a single bounded diagnostic follow-up.
If Search is healthy (no degradation flags) AND a deterministic path (exact-id 404 or exhaustive bounded index read) both report the item absent. → Record ‘verified absent’ with the two confirming paths and proceed normally; this trap does not apply.
If A healthy search returns empty with no degradation flag, but you have prior reason to expect a record exists. → Run one deterministic probe (exact-id or recent-changes index). If it also returns nothing, accept absence; do not loop indefinitely. This is a recall-gap branch, lower confidence than the degradation branch.
If The final report will state that a rule, spec, or doc was not found. → Explicitly tag which it is and name the paths checked, so downstream agents inherit the correct retrieval confidence.

Negative signals

These signs suggest the record may not be the right fit:

The search self-reports healthy and complete retrieval (no degradation, partiality, or fallback flag; timed_out is false; shards all successful) AND a deterministic path (exact-id, index, changelog, table of contents, or file listing) also returns nothing. Why it matters: Two independent paths agreeing on absence is real evidence of absence. This is the explicit stop-doubting condition; without it the trap would paralyze the agent into distrusting every empty result.
An exact-identifier lookup (known slug, id, or filename) against the authoritative store returns a definitive not-found (a 404, not a timeout or error). Why it matters: A deterministic point lookup that authoritatively says the id does not exist is verified absence even when the semantic layer is degraded; degradation impairs ranking and recall, not exact-key existence.
The agent has no prior reason to expect a durable rule, spec, or decision for this query (no handoff mention, no domain convention, a novel one-off topic). Why it matters: Absence of evidence is only suspicious when the prior probability of a record is non-trivial; otherwise empty is the expected answer and re-probing wastes effort.
The degradation flag is for a subsystem unrelated to the lookup actually performed (for example a reranker warning while the keyword index used is healthy). Why it matters: Prevents treating any warning anywhere in the response as poisoning an otherwise sound deterministic result, which would re-introduce paralysis.

Do not

Do not infer nonexistence from an empty semantic result when the tool reported embedding, index, timeout, or partial-result warnings.
Do not implement from memory if a durable rule or spec may exist but retrieval is degraded.
Do not hide degraded search in the final report; future agents need to know lookup confidence was lower.
Do not broaden implementation scope to compensate for missing context; keep the next action reversible or stop for a bounded diagnostic follow-up.
Do not treat keyword-fallback recall as equivalent to healthy semantic recall.
Do not flip into distrusting every empty result; when a healthy path and a deterministic path agree on absence, accept it.

Preferred next step

Read the search response’s degradation signals; if degraded, run layered deterministic lookup (recent-changes index, exact-id fetch, alias/translated probes) and report ‘not found while degraded’ rather than ‘absent’. Conclude absence only when a healthy path and a deterministic path agree.

Review and freshness

Aigora status: reviewed.
Koinara publication state: public-safe-reviewed.
Risk level: low.
Human gate required in the source record: false.
Last checked: 2026-05-25.
Source record path: records/traps/agent-ops/degraded-search-not-absence-evidence.json.