Control room
Aggregate telemetry only. Prompts, retrieved text, transcripts, audio, tool arguments, and reasoning are not displayed or stored here.
Recent health windows
Admission and fairness
Live aggregate reservations refreshed with this console. One active caller may use normal capacity; when multiple callers are active in the 30-second window, newly released capacity is shared. Caller IDs, API keys, prompts, sessions, and request content never appear here.
Foundation dials
Current bounded contracts, shown so operators can understand behavior before changing configuration or interpreting a benchmark.
Auth → catalog/limits → conservative prompt-work estimate → session affinity → fair admission → lane selection → Bifrost transport → stream and release.
Normalize → structure-aware chunks → exact embedding cache → Qdrant + SQLite FTS5 → RRF → rerank → opt-in evidence packing.
- 27B
- 65,536 context · 8,192 max output · one edge/provider slot per lane
- 9B
- 32,768 context · fast 4,096 output · utility 2,048 output · capacity 9
- work
- input quantum 2,048 · output quantum 1,024 · output weight 2
- overflow
- tool-free routine requests only; estimated input ≤8,192 and output ≤2,048
- fairness
- 30-second active-caller window; no waiting queue or preemption
- chunks
- target 1,050 · hard max 1,500 · overlap 100 estimated tokens
- retrieval
- dense 40 + lexical 40 → rerank 20 → return max 10
- cache
- query 128/60s · exact embeddings 100,000 · dimensions 4,096
- sessions
- opaque hash · 30-minute idle TTL · 4,096-entry bound
- HTTP
- 64 MiB body · chat 30m · default 5m · rerank 2m · ASR/TTS 10m
Traffic by channel
Lane selections
Daily usage history
Content-free aggregate totals retained across edge restarts for the last 90 days.
Alerts
Loading…
Test a chat lane
This uses the same authenticated gateway routing and limits as a client request. It is not logged as content.
API keys
Labels, application scope, policy metadata, and counters are retained; raw keys are hashed and shown only once.
| Label | Application | Policy | Prefix | Created | Requests | Status |
|---|
Routing and queue truth
The utility front door keeps routine work on the utility provider and only redirects eligible stateless requests to the locked 32B destinations. Stateful sessions and tool-bearing requests retain affinity. Each 27B provider is admitted one generation at a time; excess requests are rejected early with bounded backpressure instead of waiting indefinitely in Bifrost.
Routing decisions are persisted in daily aggregate history and exposed through the routing field of /admin/api/summary and the rassymind_routing_decisions_total Prometheus metric. No prompt, response, tool argument, retrieved text, transcript, audio, or reasoning content is stored.