Phase 3.5 extraction & assembly
Phase 3.5 — Risks and Mitigations
Risks for extraction worker + context assembly integration, including proxy hot-path latency, degradation ladder correctness, and cross-tenant isolation.
Phase 3.5 — Risks and Mitigations
| Risk | Area | Likelihood | Impact | Mitigation |
|---|---|---|---|---|
| Proxy hot-path latency regression | Context assembly + degradation ladder | Medium | High | Hard latency budgets (40ms assembly target), async/timeout wrappers, and circuit-breaker behavior verified in integration tests |
| Degradation ladder fails closed or blocks responses | Proxy integration | Low | High | Ensure assembly failure never blocks LLM responses; verify each ladder level in fixtures with real dependencies |
| Wrong/unstable extraction output schema | Worker extraction + schema v2 | Medium | High | Contract tests against extraction output shape; reject invalid payloads with safe fallbacks |
| Cross-tenant memory leakage via orchestration code | Worker + context service | Low | Critical | ISO-* cross-tenant test suite in CI; enforce org_id scope at every query path (not just at entrypoints) |
| MCP tool behavior diverges from API behavior | MCP Memory Tools | Medium | Medium | Reuse the same service layer for MCP tools and API; golden tests for search_memory/write_memory/record_feedback |
| Cost-tiered extraction execution overload | Provider-agnostic execution | Medium | Medium | Strict per-session/task timeouts; queue backpressure; bounded concurrency per cost tier |
| Dead-letter handling leads to permanent staleness | Celery tasks | Low | Medium | DLQ observability + replay path; periodic sweeper to re-drive stalled extractions |
| Context packaging correctness regressions | DP packer + formatter | Medium | Medium | Golden-output regression tests with contradiction fixtures; compare packed context to expected ordering |
Edit on GitHub
Last updated on