资产生成 Request Logs

The Logs page lists every request and asynchronous job your team sends through 资产生成, newest first. Use it to find one operation by ID, narrow the list by model or outcome, watch traffic as it arrives, and inspect routing, usage, and cost.

For aggregate charts and spend totals, see Observability. For usage grouped by model, user, or tag, see Custom Reporting.

The public REST API does not expose the Logs page's list and filter query. To look up one completed generation by ID, use GET /v1/generation or the 创意脚本's getGenerationInfo() workflow. Generation lookup returns provider, latency, token usage, cost, and finish reason.

To send routing attempts and error fields for every request to an observability backend, configure 资产生成 Trace Drains. Use the dashboard export for an ad hoc CSV or JSON file of the request rows currently loaded under your filters.

Open the Logs tab in the 资产生成 sidebar. It's available at two scopes:

• Team: every request and asynchronous job across your team, at /[team]/~/ai-gateway/logs
• Project: only that project's logs, at /[team]/[project]/ai-gateway/logs

The two views behave identically. The project view is the team view with a project filter applied.

Each row is one request or asynchronous job. The table scrolls horizontally when the available width can't fit every column, including when request details are open.

Usage holds one slot per row and changes what it reports based on the model's modality:

Hover a Usage cell for the long form, which names the components. For language models that means the input and cache-read halves of the left figure, and the reasoning and output halves of the right.

A dash means the request reported nothing for that column. A request that failed before reaching a provider shows one for both Usage and Cost.

Running asynchronous jobs appear first. The remaining rows are ordered newest first, and the list can't be re-sorted. To narrow it, use the filters below.

The filter bar sits above the table.

Search matches the model, provider, or request ID. Enter an exact job ID to find an asynchronous job.

Dropdowns each accept multiple values:

Date range offers presets from the last 5 minutes to the last 30 days. You can look back at most 36 days.

Every filter is stored in the URL, so you can share a filtered view by copying the address. Reset clears the search and filters.

The chart above the table shows token volume under the same filters. When matching operations report no token usage, the chart shows request volume instead. If the range contains no requests, the chart shows an empty state. Drag across a populated chart to select that time range and zoom into a finer granularity.

Turn on Live to tail logs as they arrive. New rows appear at the top, refreshing every 5 seconds.

Live mode has three limits worth knowing:

• It only works with a relative range. Turning it on while an absolute range is selected switches you to the last hour.
• Scrolling back through history is disabled while live. Pause to load older logs.
• Requests take about 90 seconds to fully ingest, so the newest rows lag real time by roughly that much.

With Live off, the list loads 50 more rows each time you scroll to the bottom.

Click any row to open its details beside the list. The panel is resizable by dragging its left edge. The selected log ID is stored in the URL, so you can copy the address to share the same log and filtered list.

The header carries the generation or job ID with a copy button. Request started identifies the authentication input, inference region, and whether Zero Data Retention applied. OIDC-authenticated requests name their project because the token itself has no API key name. App tokens and personal access tokens appear as team-scoped authentication.

Routing describes what happened after 资产生成 received the request. A single provider attempt appears in one Routing card. A request that needed multiple attempts shows one Routed to [provider] card per attempt. Each card can include the provider, provider region, served model, credential source, time to first token, compact timing spans, status, and provider response.

Timing spans use green for successful work, amber for a 4xx response, and red for a 5xx response or timeout. A recovered request also has a refresh icon in the table's Status cell.

Usage and Cost break the request down line by line, so you can see which tokens and which charges made up the total.

Use the Request Mode filter to show asynchronous jobs, or search for an exact job ID. Click a job to open its details beside the list:

Running asynchronous jobs stay at the top of the table so you can monitor them until they complete, fail, or are canceled. The timeline identifies when 资产生成 submitted the job. The details include its authentication, model, provider, and inference region.

Copy visible logs copies every loaded row, including job-only rows, to your clipboard. Export to CSV and Export to JSON include loaded request rows under the active filters and time range. The menu discloses job-only rows that the file export omits because those jobs don't have a flat request record yet.

Routing attempt details are kept for 30 days. Older requests still appear in the list but open with a message saying their routing details are no longer available.

Viewing logs requires read access to your team's usage. Team roles without it see a permission message in place of the table rather than an empty list.