Observability

Request Logs

Every request through VectorAxis (chat, prompts, and media) is captured as a structured log with the who, what, how-much, and how-fast: provider, model, tokens, cost, latency, cache status, retries, fallback, and guardrail outcome. It’s your single place to debug a bad response, attribute spend, and audit what happened.

Why it matters

Debug fast. Open any request to see the exact model, timing, error, and (optionally) the full request and response bodies.
Attribute cost. Every log carries its dollar cost and the key/config/user behind it, so you know who spent what.
Audit & compliance. A durable record of every call, including which guardrail ran and how it resolved.

What’s captured on every log

Field	Meaning
request_id	Unique id for the request (also returned to the caller for correlation).
created_at	When the request was served.
provider / model	Which provider and model actually served it.
status_code	HTTP result (200, 4xx, 5xx).
duration_ms	End-to-end latency in milliseconds.
prompt_tokens / completion_tokens / total_tokens	Token counts (populated for streaming too).
cost	Computed cost in USD.
cache_status	HIT / SEMANTIC_HIT / MISS / REFRESH / BYPASS / DISABLED.
retry_count / fallback_provider	Retries performed and the backup provider used, if any.
virtual_key_slug / config_slug	The virtual key and/or routing config the request used.
guardrail_slug / guardrail_status	The guardrail applied and its outcome (PASS / BLOCK / REDACT / MONITOR / BYPASS).
organization / workspace_id / user_id	Attribution: who the request belongs to.

The detail view adds more. Opening a single log also shows error_message and, when enabled, the full request_body and response_body. For media requests it includes operation_type plus format details: file_size_bytes, mime_type, media_format, image_size / image_quality / image_count, audio_duration_seconds, and voice. For System One decisions (operation_type system_one) it shows resolved_model (the version behind an alias like jev-latest), question_count, and min_confidence, the least certain choice or score answer in the call.

Attribution depends on how you authenticate. Every request is logged with token usage and computed cost either way. But the attribution fields (organization, workspace_id, and virtual_key_slug) are only populated automatically when you call with a virtual key (its owning org and workspace are authoritative). With raw direct credentials (x-provider + x-api-key) those fields stay empty unless you pass x-organization / x-workspace-id yourself, so per-org, per-workspace, and per-key filtering and roll-ups won’t find the request.

Searching & filtering

The Logs page (and GET /v1/logs) let you narrow to exactly the requests you care about, then click a row for the full detail. Every filter is a query parameter:

Filter	Narrows to…
provider, model	a specific provider and/or model.
status	success or error.
cache_status	hits, misses, bypass, etc.
virtual_key_slug, config_slug, user_id	a specific key, routing config, or user.
guardrail_slug, guardrail_status	requests a guardrail touched, and how it resolved.
operation_type	one kind of operation, e.g. system_one or image_generation. Chat completions have none, so setting it excludes them.
start_time, end_time	a date/time range.
min_cost/max_cost, min_duration_ms/max_duration_ms, min_tokens/max_tokens	expensive, slow, or large requests.
search	free-text across the request.
request_id	one exact request.
sort_by / sort_dir, page / page_size	ordering and paging.

Scope any query to a workspace with workspace_id. An org member must scope to a workspace they belong to; org admins can read across the whole org.

Export

Export a filtered set of logs as CSV or JSON (up to 50,000 rows). Because exports can be large, they run asynchronously: you start a job, poll it until it’s DONE, then download the file. Finished exports are available for 1 hour. Exporting is an org-admin action.

Who can view logs

Workspace members and above can view and filter logs for their workspace; org admins see the whole org. Exporting requires an org admin. Full breakdown on the Access Control page.

Developer API reference

Method	Path	Who can call it
GET	/v1/logs	Member (scope to a workspace)
GET	/v1/logs/{id}	Workspace member
POST	/v1/logs/export	Org admin
GET	/v1/logs/export/{jobId}	Org admin
GET	/v1/logs/export/{jobId}/download	Org admin

http

# Slow, expensive OpenAI errors in a workspace, last 24h
GET /v1/logs?organization=acme&workspace_id=<uuid>&provider=OPENAI
    &status=error&min_duration_ms=5000&start_time=2026-06-29T00:00:00Z
Authorization: Bearer va_...

# Kick off an export, then download when DONE
POST /v1/logs/export        { "format": "csv", "provider": "OPENAI" }   -> { "job_id": "...", "status": "PENDING" }
GET  /v1/logs/export/{jobId}                                            -> { "status": "DONE", "row_count": 1234 }
GET  /v1/logs/export/{jobId}/download                                   -> file

Want the aggregate picture instead of individual rows (cost and latency trends, top models, cache-hit rate)? See Analytics.

VectorAxis is built by GK Labs, a member of NVIDIA Inception.