What happened to this call?
Search by request ID and inspect the requested model, resolved route, provider, region, status, token counts, cost, cache state, and timing.
Request-level evidence
LLM observability from request to settlement
When an AI request is slow, costly, or fails, a model name is not enough. LLMRegistry connects each request ID to its route, provider, token use, cost, timing, cache result, and final outcome.
- Core record
- One settled request ledger
- Useful timing
- Start time and total response time
- Content boundary
- Metadata without prompts
Answer the questions that appear during an incident
The monitoring view uses completed request and provider evidence. Empty periods stay empty instead of being presented as perfect delivery.
Is one provider struggling?
Compare observed request counts, delivery percentage, time to first token, fallback count, traffic share, and current circuit state.
Where did the money go?
Connect provider cost, platform charges, cache outcomes, model races, and route choices to one request instead of comparing separate reports.
How the evidence stays useful
Each chart and detail view should name the thing it actually measured.
Give every request an ID
The gateway returns a request identifier that can connect an application error to the matching ledger row and provider result.
Record the resolved path
The ledger shows the route, model, provider mode, region, fallback count, and cache or race result that produced the final outcome.
Separate speed measures
Time to first token shows when output began. Total response time shows the full gateway round trip. The interface does not relabel one as the other.
Keep unknown values honest
If a signal is not measured or configured, LLMRegistry shows it as unknown or unavailable instead of creating sample success data.
Operational evidence without a prompt archive
Troubleshooting usually needs the path and outcome, not the customer’s exact words.
Recorded metadata
- Request ID, time, endpoint, project, route, model, and provider.
- Prompt and completion token counts.
- Provider cost, billed cost, cache state, and fallback result.
- Response timing, status, circuit events, and selected region.
Not presented as measured
- Prompt or completion text in normal usage reporting.
- Gateway-only overhead when only total response time exists.
- Alert or webhook history until those features have durable records.
- Success percentages for a period with no completed traffic.
Related controls
Make each AI request explainable
Use one request ID to follow routing, provider work, usage, and cost without turning monitoring into a content store.
Get started