Open Questions and Risks¶
Questions¶
- Which Analyst API repository/version is the structured engine in production?
- Are project IDs enforced by every backend service and index?
- What are the source authority and update cadence for tariff schedules?
- Which LLM/provider is used for each use case and fallback?
- How are Qdrant, Meilisearch, MinIO, and Postgres backed up and reindexed?
- What are target answer citation precision, latency, and cost metrics?
Risks¶
| Risk | Mitigation |
|---|---|
| Private corpus leakage | Cross-project authorization tests and index filters |
| Uncited fallback answers | Label fallback confidence and require review |
| Parser failures | Quarantine/retry ingestion and retain source files |
| Index drift | Version embeddings and reconcile all stores |
| LLM cost/latency | Cache, budgets, model tiers, and timeouts |
| Shared SSO secret compromise | Rotation and short-lived tokens |
Verification checklist¶
- Run complete upload → index → hybrid query → citation flow.
- Query each use-case project with adversarial cross-project terms.
- Validate private dossier cannot be selected or retrieved publicly.
- Exercise Celery failure/retry and filesystem sync recovery.
- Measure retrieval stage timings and source precision.