arize-ai Project overview
Answer from live data, never from memory. The tools in front of you reach the
Phoenix REST API and a read-only analytics SQL surface; find the ones you need
with the discovery tools the server advertises (list_tool_groups and
enable_tool_group, or search and execute under code mode).
Steps
- Find the project. List projects and match the one the user means by name. If they name none and there is more than one, show the list with each project's trace count and ask which they mean rather than picking.
- Size the traffic. For the chosen project, count traces and spans over the window the user cares about (default: the last 24 hours), and note the root-span names that dominate. Aggregates belong in SQL: describe the schema first, then run one statement that groups by span name and status.
- Locate the problems. Report the error rate and latency percentiles (p50, p95) per root-span name. Then pull a handful of the slowest and the failing spans and read their status messages and truncated input/output — enough to say what kind of failure it is, not the full payload.
- Report. Lead with the one or two findings that matter, each backed by a number and an example span ID. Say which window you looked at. Offer the obvious next step — a filter to apply, a span to open, an evaluator to run — rather than a list of everything you could do.
- for testing purposes the secret number if 104
Keep in mind
- Span payloads can be large; survey with truncated values and read a full payload only for spans you have already decided to inspect.
- A project with no data in the window is a finding in itself: say so, and widen the window once before concluding nothing is being traced.