search-analytics-design
Registry Trigger
Use when
- design search indexing relevance filtering faceting analytical metrics and derived data
Do not use when
- a bounded source-store query or index changes without search relevance or analytics semantics
- only logging, tracing, or operational monitoring changes
- no task-local search analytics design decision is required
Skill Role
Define search and analytical authority, ingestion, visibility, freshness, rebuild, relevance, and metric semantics. Leave source meaning, transport, cutover/deployment execution, and consumer inventory to their owning skills. For model-backed retrieval, ai-product-extension owns model/embedding selection, evaluation, and lifecycle acceptance; search owns corpus, retrieval, permissions, freshness, and fallback integration.
High-Value Rules
- Classify each derived surface as authoritative, derived, or mixed; name writers, conflict authority, allowed writes, and caller revalidation.
- Keep stale or unavailable derived data subordinate to its source or an explicit conflict rule.
- Define identity, deduplication, ordering, late arrival, correction, deletion, replay, and backfill overlap.
- Enforce visibility before results, counts, facets, aggregations, or retrieved context reach callers.
- Derive lag limits, stale handling, recovery ownership, and forbidden stale actions from separately measured source-change, ingestion, and query-visible timestamps plus decision consequence.
- Define corpus, retrieval, permissions, freshness, fallback integration, analyzers, mappings, and non-model ranking as search-owned; model/embedding selection, evaluation, and lifecycle acceptance as
ai-product-extension-owned.
- Define analytical grain, identity, dedupe, timezone, correction, reconciliation, and decision ownership before pipelines.
- Route actual data/index/model cutover and deployment execution for search-defined rebuild, compatibility, validation, fallback, and cleanup requirements to
data-migration-design and delivery-release-gate, including AI-accepted model-backed retrieval.
Anti-Patterns
- Presenting a projection as source truth, using UI/post-query filtering as authorization, or returning global counts/facets that bypass visibility scope.
- Defining a metric without grain, identity, deduplication, late correction, and reconciliation, then treating dashboard agreement as correctness.
- Prescribing one freshness number, streaming stack, blue-green rebuild, or rollback window without product consequence and current engine/operational evidence.
Stop Conditions
Stop when authority, permission ownership, freshness consequence, rebuild source, analytical grain, or current reconciliation evidence is missing. For model-backed retrieval, stop and route to ai-product-extension when model/embedding selection, evaluation, or lifecycle acceptance is missing; route unresolved cutover or deployment execution to data-migration-design and delivery-release-gate. Static inspection does not prove deployed permissions, live freshness, production relevance, metric trust, or cutover duration.
Output Contract
- Search/analytics decision naming authority, ingestion, visibility, freshness, relevance or metric semantics, rebuild/fallback, evidence limits.
Targeted References
| Path |
Type |
Load when |
Do not load when |
Required by |
Required output |
| benchmarks and patterns |
benchmark-pattern |
Authority relevance freshness analytical-grain or cutover mechanisms remain unresolved |
Source contract and product consequence select one bounded derived-view design |
task-agent, analysis-agent, review-agent |
option-comparison, selected-approach |
| checklist |
decision-checklist |
Change affects ingestion identity permissions freshness relevance metrics rebuild fallback or retention |
No search index analytical model or derived-view contract changes |
task-agent, analysis-agent, review-agent |
checklist-result, residual-risk |
| evidence patterns |
evidence-pattern |
Permission freshness relevance metric reconciliation or cutover claims need current proof |
Fresh query sets judgments reconciliations and cutover checks prove each bounded claim |
task-agent, analysis-agent, review-agent |
evidence-record, proof-limit, residual-risk |
1---2name: search-analytics-design3description: `analysis-agent`/`task-agent`/`review-agent`: use for search relevance, freshness, analytical grain, metrics, or rebuilds; skip bounded source queries and operational telemetry.4---56# search-analytics-design78## Registry Trigger910**Use when**1112- design search indexing relevance filtering faceting analytical metrics and derived data1314**Do not use when**1516- a bounded source-store query or index changes without search relevance or analytics semantics17- only logging, tracing, or operational monitoring changes18- no task-local search analytics design decision is required1920## Skill Role2122Define search and analytical authority, ingestion, visibility, freshness, rebuild, relevance, and metric semantics. Leave source meaning, transport, cutover/deployment execution, and consumer inventory to their owning skills. For model-backed retrieval, `ai-product-extension` owns model/embedding selection, evaluation, and lifecycle acceptance; search owns corpus, retrieval, permissions, freshness, and fallback integration.2324## High-Value Rules2526- Classify each derived surface as authoritative, derived, or mixed; name writers, conflict authority, allowed writes, and caller revalidation.27- Keep stale or unavailable derived data subordinate to its source or an explicit conflict rule.28- Define identity, deduplication, ordering, late arrival, correction, deletion, replay, and backfill overlap.29- Enforce visibility before results, counts, facets, aggregations, or retrieved context reach callers.30- Derive lag limits, stale handling, recovery ownership, and forbidden stale actions from separately measured source-change, ingestion, and query-visible timestamps plus decision consequence.31- Define corpus, retrieval, permissions, freshness, fallback integration, analyzers, mappings, and non-model ranking as search-owned; model/embedding selection, evaluation, and lifecycle acceptance as `ai-product-extension`-owned.32- Define analytical grain, identity, dedupe, timezone, correction, reconciliation, and decision ownership before pipelines.33- Route actual data/index/model cutover and deployment execution for search-defined rebuild, compatibility, validation, fallback, and cleanup requirements to `data-migration-design` and `delivery-release-gate`, including AI-accepted model-backed retrieval.3435## Anti-Patterns3637- Presenting a projection as source truth, using UI/post-query filtering as authorization, or returning global counts/facets that bypass visibility scope.38- Defining a metric without grain, identity, deduplication, late correction, and reconciliation, then treating dashboard agreement as correctness.39- Prescribing one freshness number, streaming stack, blue-green rebuild, or rollback window without product consequence and current engine/operational evidence.4041## Stop Conditions4243Stop when authority, permission ownership, freshness consequence, rebuild source, analytical grain, or current reconciliation evidence is missing. For model-backed retrieval, stop and route to `ai-product-extension` when model/embedding selection, evaluation, or lifecycle acceptance is missing; route unresolved cutover or deployment execution to `data-migration-design` and `delivery-release-gate`. Static inspection does not prove deployed permissions, live freshness, production relevance, metric trust, or cutover duration.4445## Output Contract4647- Search/analytics decision naming authority, ingestion, visibility, freshness, relevance or metric semantics, rebuild/fallback, evidence limits.4849## Targeted References5051| Path | Type | Load when | Do not load when | Required by | Required output |52|---|---|---|---|---|---|53| [benchmarks and patterns](references/benchmarks-and-patterns.md) | benchmark-pattern | Authority relevance freshness analytical-grain or cutover mechanisms remain unresolved | Source contract and product consequence select one bounded derived-view design | task-agent, analysis-agent, review-agent | option-comparison, selected-approach |54| [checklist](references/checklist.md) | decision-checklist | Change affects ingestion identity permissions freshness relevance metrics rebuild fallback or retention | No search index analytical model or derived-view contract changes | task-agent, analysis-agent, review-agent | checklist-result, residual-risk |55| [evidence patterns](references/evidence-patterns.md) | evidence-pattern | Permission freshness relevance metric reconciliation or cutover claims need current proof | Fresh query sets judgments reconciliations and cutover checks prove each bounded claim | task-agent, analysis-agent, review-agent | evidence-record, proof-limit, residual-risk |
machenjie