A useful gateway log lets an investigator reconstruct who was authorized, which request attempts occurred, and what usage was recorded. Start with a field contract and correlation rules. A large log archive is not automatically sufficient evidence.
Keep identity, attempts, and usage distinct #
| Field group | Recommended information | Validation check |
|---|---|---|
| Event identity | Event ID, logical request ID, attempt ID, UTC time, emitter | Can retries be separated from duplicate delivery? |
| Access identity | Pseudonymous credential and workload references | Is identity authenticated or merely client-supplied? |
| Routing | Model, provider, route, deployment reference | Can fallback to a different model be reconstructed? |
| Usage | Input/output tokens, cost value, currency and calculation basis | Are unavailable values null rather than zero? |
| Outcome | Success/failure, numeric status where known, duration | Are cancellations and streaming completion handled? |
| Context | Trusted network reference, deployment or job reference | Does the field identify the client or just the gateway? |
An illustrative event #
This is a proposed investigation format, not the current InferTrail ingestion schema. All identifiers and numbers below are synthetic.
{
"event_id": "evt-example-1",
"logical_request_id": "req-example-1",
"attempt_id": "attempt-2",
"timestamp": "2026-09-30T12:00:00Z",
"principal_id": "hmac-key-example",
"workload_id": null,
"model": "deployment-a",
"input_tokens": 200,
"output_tokens": 80,
"estimated_cost_usd": null,
"outcome": "success"
}
A missing workload ID means the request cannot yet be assigned to an originating job. It does not mean every such request belongs to the same anonymous user. Preserve that distinction during aggregation.
Choose an explicit content boundary #
Allowlist fields before export. Exclude raw credentials, authorization headers, cookies, prompts, responses, and free-text error bodies from the metadata dataset. Test unexpected nested fields and failure paths, not only a clean success example. Restrict raw lookup mappings and keep tenant identifiers separated.
Check LiteLLM's actual emitted payload #
LiteLLM documents a standard logging payload with usage, timing, and identity metadata. Optional identity can be absent, and some callback events lack the standard payload. Validate your deployed version and success, failure, streaming, and retry paths against the official specification; do not assume every documented field is present on every event.
Measure loss and retention #
Track accepted, written, rejected, and dropped events at each collector stage. Reconcile totals against a known request sample. Test duplicate delivery, queue overflow, a full disk, and shutdown. A collector that silently loses busy periods creates exactly the gap an investigator needs to understand.
Set retention from investigation requirements and applicable organizational policy. Verify that the lookup mapping and request records overlap in time; retaining one without the other can make attribution impossible. Document deletion and access controls along with the event schema.
Next: correlate gateway hops and use the records in a key-abuse investigation.