The LLM Calls page is the single place to find any LLM invocation that has ever passed through Observatory. Use it when you know what you’re looking for — a user, an agent, an error spike, a slow call — and you want to drill into the actual prompt and completion.Documentation Index
Fetch the complete documentation index at: https://docs.flowx.ai/llms.txt
Use this file to discover all available pages before exploring further.
Key features
Lazy-loaded table
Server-side pagination, sorting, and filtering. The page stays fast on datasets of millions of rows.
Global text search
The search box runs across agent name, model, input, and output.
Filter by agent
Quick filter to scope the table to a single agent name.
Time range
Last Hour, 3H, 24H, 7D, 1M, 6M selector at the page header level.
Status badges
Success, error, in-progress, and timed-out are colour-coded at a glance.
Detail drawer
Click any row to open the full input, output, metadata, and cost breakdown.
Columns
| Column | What it shows |
|---|---|
| Status | success, error, started, or timed-out (started > 5 min). |
| Start time | When the call began, in the viewer’s local timezone. |
| Agent | The @agent("name") decorator value. |
| Model | Provider model identifier — gpt-4, claude-opus-4-7, etc. |
| Duration | Wall-clock seconds. Below 4s is success, 4–8s warning, >8s danger. |
| User | The user who triggered the call, when available in event metadata. |
| Tokens | Total tokens for the call (prompt + completion). |
| Requests | Number of HTTP requests bundled into the same logical call. |
Filtering and search
The text search box matchescontains against the agent name, model, input, and output fields. The agent dropdown is the equivalent of an exact-match filter on the Agent column.
For more complex queries — status, model, tags, latency bounds, cost — use the API directly:
The detail drawer
Click any row to open the run detail drawer. It contains:- The full prompt and completion (rendered as text or JSON depending on payload)
- The chain-of-calls tree if the run contained nested chains or tools — links to Traces
- Token and cost breakdown by model
- Metadata extracted by the SDK (
user_email, custom tags, request ID)
Common workflows
Find every failed call for one agent in the last 24 hours
Find every failed call for one agent in the last 24 hours
- Set the time range to 24H.
- Pick the agent from the Filter by agent dropdown.
- Sort by Status to bring errors to the top.
Find calls slower than 8 seconds
Find calls slower than 8 seconds
Sort by Duration descending. Rows with the red
danger badge are above the 8s threshold.Find a call by something the user typed
Find a call by something the user typed
Paste a snippet of the user’s input into the search box. Matching rows surface immediately.
Related resources
Traces
For runs that contain nested chains and tools, see the trace view.
Analytics
Aggregate trends across the same data.