Intelligence
Asynchronous workers, extraction, embeddings, deletion orchestration, and billing rollups.
The intelligence layer runs work that must not block the proxy: session extraction, memory writes, embeddings, organization deletion, billing reconciliation, maintenance, and MCP audit fan-out.
Provisional Ledger class: provisional
Verified 2026-10-21Several tasks are implemented; queue durability and exactly-once projection are not certified
Pages
- Workers — queues, task families, dead letters, and enqueue admission.
- Extraction — provider adapters, quality evaluation, and manual vLLM limits.
- Embedder — CPU stub, TEI, hosted backends, and readiness geometry.
- Extraction evaluation — gold set, cassette replay, contract gates, and regression limits.
What is implemented
| Capability | Status | Boundary |
|---|---|---|
| Session extraction and memory writes | IMPLEMENTED | Cassette and unit evidence; live vLLM is manual |
| OpenAI and vLLM extraction providers | IMPLEMENTED | vLLM has no CI GPU runner |
Embedding profiles (cpu, gpu, hosted) | IMPLEMENTED | CPU stub is not a quality model |
| Organization deletion orchestration | IMPLEMENTED | Full deletion certification remains open |
| Billing reconciliation and rollups | IMPLEMENTED | Hosted reconciliation evidence open |
| Dead-letter persistence, metrics, OTel | IMPLEMENTED | Alert rules exist; runbook practice is separate |
| Scheduled maintenance sweep | DEFERRED | Registered as a no-op, not business logic |
Queue layout
| Queue | Purpose | Priority |
|---|---|---|
extraction | Session turns to memories | First |
embedding | Async vector work | Second |
maintenance | Beat sweeps and housekeeping | Third |
mcp_audit | MCP tool-call audit fan-out | Last |
Redis DB allocation is deliberate: cache 0, Celery broker 1, rate limiting 2, results 3. Do not collapse them into one database.
Non-negotiable boundary
A queued task is not a completed effect. Extraction can be retried, enqueue can be idempotent, and dead letters can be persisted, but the system does not currently claim exactly-once projection into every downstream store. Read Evidence and relay before treating an async side effect as durable.