Knowledge Platform — Observability¶
Target
This page describes intended behaviour or structure. No implementation is claimed unless a section is explicitly marked Current implementation or Evidence.
Spans¶
Context assembly and retrieval spans join the run trace (run.id, attempt.id) with context.snapshot.id, retrieval.query.id, the index revisions used and the counts per section; embedding calls carry the GenAI attributes from the model gateway. Ingestion spans are keyed on source.id and snapshot.id with one child per stage. Content never appears on spans.
Metrics¶
| Metric | Why it matters |
|---|---|
connectsoft.factory.knowledge.sources by state |
stale and failing sources |
connectsoft.factory.knowledge.snapshots, ingestion_stage_duration |
pipeline throughput and slow stages |
connectsoft.factory.knowledge.host.index_lag |
freshness of what retrieval can see |
connectsoft.factory.knowledge.retrievals, retrieval_latency |
retrieval reliability and the 2 s interactive target |
connectsoft.factory.knowledge.index_degraded |
reduced recall in progress |
connectsoft.factory.knowledge.graph_facts by trust tier |
growth of inferred versus source facts |
connectsoft.factory.context.assemblies by status, tokens_used, dropped_items |
how often budgets truncate and why |
connectsoft.factory.artifacts.stored, bytes, deleted |
storage meter and lifecycle |
connectsoft.factory.memory.items, in_review, promotions, skill_candidates |
learning loop health and review backlog |
connectsoft.factory.repo_intelligence.index_builds, tool_calls |
repository index freshness and tool usage |
connectsoft.factory.knowledge.deletion_pending |
data lifecycle debt |
Logs¶
Structured, keyed on source, snapshot, run and attempt ids; no content (D31). Retention 30 days.
Health¶
/knowledge/health reports the knowledge store, vector index, blob store, model gateway reachability for embeddings and connector health; readiness fails when the vector index is unreachable and no full-text fallback exists for a mandatory collection.
Objectives (pilot class, D30)¶
| Objective | Target |
|---|---|
| Retrieval latency | p95 under 1 s inside the 2 s interactive budget |
| Context assembly | p95 under 3 s for the default budget |
| Index lag after publication | under 15 minutes |
| Staleness detection | within the configured window (72 h default) |
| Deletion lineage | executed within 30 days, evidence written |
Dashboards and alerts¶
Dashboards: source freshness, pipeline stages, index revisions and lag, retrieval latency and degradation, context budgets and truncation, memory review backlog, artifact storage per tenant, deletion debt. Alerts: sources Failing, index lag above threshold, index degraded, context Failed rate, review backlog age, deletion items past the window.