Filter-Based Annotation Delete Endpoints#
May 5, 2026
Available in arize-phoenix 15.4.0+
Three new DELETE endpoints let you bulk-remove annotations from a project by filter — without knowing every individual span/trace/session ID. This closes the annotation lifecycle loop for automated pipelines that tag annotations with a custom identifier on creation and need to roll them back later.
Query parameters (all optional, but at least one time-bound or delete_all=true is required):
| Parameter | Description |
|---|---|
name | Exact match on annotation name |
identifier | Exact match on annotation identifier |
annotator_kind | LLM, CODE, or HUMAN |
start_time | Inclusive lower bound on created_at |
end_time | Exclusive upper bound on created_at |
delete_all | Set true to waive the time-bound requirement |
Token Counts in Trace and Session REST Payloads#
May 5, 2026
Available in arize-phoenix 15.4.0+
The GET /v1/projects/{project_identifier}/traces and GET /v1/projects/{project_identifier}/sessions endpoints now include cumulative token usage fields — cumulative_token_count_prompt, cumulative_token_count_completion, and cumulative_token_count_total — so you can read aggregate token consumption directly from the REST API without recomputing from raw span attributes.
Values are summed from root spans and default to 0 for traces or sessions with no LLM calls. The /v1/spans endpoint is unchanged — span-level token counts remain in the existing attributes dictionary.
Experiment CSV Export Includes Dataset Metadata#
May 5, 2026
Available in arize-phoenix 15.3.0+
Downloading an experiment as CSV now includes per-example dataset metadata columns. Each metadata key appears as a metadata_<key> column — matching the format used by the dataset CSV export — so you can cross-reference experiment results with the original dataset context without a separate download.
Evals: Runtime Model Capability Detection#
May 5, 2026
Available in arize-phoenix-evals 3.1.0+
The OpenAI evaluator adapter now detects structured-output and tool-call support at runtime rather than checking against a hardcoded model list. This unblocks OpenAI reasoning models (o1, o3, o3-mini, o4-mini) for use with ClassificationEvaluator and ensures new models work automatically without requiring a library update.
The adapter tries structured output first, falls back to tool calling if unsupported, and caches the result per adapter instance — matching the approach already used by the Google GenAI adapter.