This release helps you move from question to action faster. PXI can complete multi-step workflows across the Phoenix UI, a new evaluator measures whether retrieved information is useful, and new trace and REST tools reduce repetitive work.

Complete Multi-Step Workflows with PXI#

August 25, 2026

Available in arize-phoenix 20.4.0+

PXI can now discover the actions available on the current page and combine them into a single workflow. For example, it can select a model, run the playground, and inspect the output in one turn.

• Work across the UI in fewer turns — PXI can read page state, make decisions, and perform several actions as one workflow.
• Stay in control of GraphQL changes — in manual edit-permission mode, Phoenix shows each proposed mutation with Accept and Reject controls before applying it. Approval requests remain available after a page reload.
• Navigate with approval — PXI asks before moving to another page and stops if you reject the request.
• Query Phoenix data directly — PXI can use Phoenix’s read-only REST and analytics SQL tools without a network round trip or separate MCP configuration.
• Recover from unavailable actions — when an action is not available on the current page, PXI receives guidance about where it can run.

Evaluate Retrieval Relevance Across Any Source#

August 21, 2026

Available in arize-phoenix-evals 3.5.0+ (Python) and @arizeai/phoenix-evals 2.3.0+ (TypeScript)

Use the new Retrieval Relevance evaluator to measure whether retrieved information helps answer the request it was meant to serve. It evaluates the combined result of each retrieval step, whether the content came from vector search, a tool or MCP call, web search, or a SQL query.

The evaluator returns a relevant or irrelevant label, a numeric score, and the judge’s explanation.

from phoenix.evals import LLMfrom phoenix.evals.metrics import RetrievalRelevanceEvaluatorllm = LLM(provider="openai", model="gpt-4o")evaluator = RetrievalRelevanceEvaluator(llm=llm)scores = evaluator.evaluate(    {        "input": "What is the return policy?",        "context": "Customers can return purchases within 30 days.",    })print(scores[0])

import { openai } from "@ai-sdk/openai";import { createRetrievalRelevanceEvaluator } from "@arizeai/phoenix-evals";const evaluator = createRetrievalRelevanceEvaluator({  model: openai("gpt-4o"),});const result = await evaluator.evaluate({  input: "What is the return policy?",  context: "Customers can return purchases within 30 days.",});console.log(result.label);

Build Focused Trace Views Faster#

August 25, 2026

Available in arize-phoenix 20.4.0+

New filtering and table controls make it easier to move from a broad project view to the traces that need attention.

• Write trace filters in plain English — toggle AI query in the trace filter field, describe the traces you want, and press Enter to generate a validated filter expression. Press Escape to restore the previous expression.
• Zoom into a chart interval with one click — click a project metric chart bin to apply its exact time range to the page, or drag across the chart to select a wider interval.
• Compare annotations in dedicated columns — show, hide, and reorder individual annotation columns on the spans, traces, and sessions tables. Hover for details or add a matching filter from the cell popover.

Manage Retention and Model Providers over REST#

August 25, 2026

Available in arize-phoenix 20.4.0+

The v1 REST API now supports project retention assignments and model-provider discovery.

• PATCH /v1/projects/{project_identifier}/retention — assign an existing trace retention policy to a project or reset the project to the default policy. This admin-only endpoint changes the assignment; it does not create or modify policies.
• GET /v1/model_providers — list the built-in provider families enabled for the deployment.
• GET /v1/custom_model_providers — list user-defined providers with cursor pagination. Phoenix never returns encrypted provider credentials.

Assign a retention policy:

curl -X PATCH "$PHOENIX_ENDPOINT/v1/projects/my-project/retention" \  -H "Authorization: Bearer $PHOENIX_API_KEY" \  -H "Content-Type: application/json" \  -d '{"policy_id": "UHJvamVjdFRyYWNlUmV0ZW50aW9uUG9saWN5OjI="}'

Reset a project to the default policy by sending {"policy_id": null}.

List the built-in model providers:

curl "$PHOENIX_ENDPOINT/v1/model_providers" \  -H "Authorization: Bearer $PHOENIX_API_KEY"

Additional Improvements#

August 20–25, 2026

Available in arize-phoenix 20.4.0+ and arize-phoenix-evals 3.5.1+

• Use native scrollbars — enable your browser and operating system’s scrollbar style from Profile → Accessibility.
• Get clearer not-found pages — missing resources and unmatched URLs now show a useful 404 page instead of a blank page or generic error.
• Submit login forms with Enter without navigating to the forgot-password page.
• Read long filters and share tooltips more easily without content overflowing clipped layouts.
• Track current model costs with refreshed built-in token prices and preserved prompt-cache usage from the OpenAI-compatible /v1/chat/completions proxy.
• Use Anthropic SDK v1 with Phoenix Evals by upgrading anthropic alongside arize-phoenix-evals 3.5.1+.