​

OpenAI-Compatible Chat Completions Endpoint

• Model IDs name the provider — {provider}:{model_name} for a built-in provider (openai:gpt-4o, anthropic:claude-sonnet-4-5), or custom:{provider_id}:{model_name} for a custom provider record stored in Phoenix. Only the first colons are split, so model names that contain colons survive intact.
• Streaming — set stream: true for server-sent chat.completion.chunk events terminated by data: [DONE]. stream_options: {include_usage: true} appends a final usage chunk.
• Familiar parameters — temperature, top_p, max_tokens / max_completion_tokens, stop, seed, frequency_penalty, and presence_penalty all pass through.
• OpenAI-shaped errors — every failure, including validation errors, comes back as {"error": {"message", "type", "code"}}. Provider HTTP errors forward their status; an unreachable provider becomes a 502.
• Available to every authenticated role — the endpoint writes nothing, so viewers can call it.

curl -X POST "$PHOENIX_HOST/v1/chat/completions" \
  -H "Authorization: Bearer $PHOENIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai:gpt-4o",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Say hello."}
    ]
  }'

import os

from openai import OpenAI

client = OpenAI(
    base_url=f"{os.environ['PHOENIX_HOST']}/v1",
    api_key=os.environ["PHOENIX_API_KEY"],
)

completion = client.chat.completions.create(
    model="anthropic:claude-sonnet-4-5",
    messages=[{"role": "user", "content": "Summarize what a span is in one sentence."}],
)
print(completion.choices[0].message.content)

​

AI Query for Filter Fields

• Enter converts — in plain-English mode, Enter translates your prose into a filter expression, streaming it in as it forms. From expression mode, ⌘/Ctrl+Enter hands the current draft to AI query directly.
• Validated, with one correction round — the generated expression goes through the same validator the field itself uses, and the model gets one chance to fix a rejected expression before you see it.
• Escape undoes — Escape walks back whatever AI query last did, restoring your original phrasing.
• The model’s vocabulary is the field’s vocabulary — the field names and examples handed to the model are derived from the same completions and snippets that power the typeahead, so the two can’t drift apart.

• Browser AI — the browser’s built-in on-device model (Chrome and Edge’s Prompt API). No credentials, no network round trip, and the default wherever a built-in model is available. A companion card shows download status and can fetch the model ahead of first use.
• Any provider Phoenix knows — built-in providers, Azure, Bedrock, and stored custom providers, all called through the new /v1/chat/completions proxy so no API key ever reaches the browser.

​

Annotation Metric Charts and Uncapped Chart Selection

• A chart per annotation name — the project Metrics page adds span, trace, and session annotation sections, each followed by a grid with one chart per annotation name on that level. Every chart plots mean score over time, with a score/label toggle for annotations that carry both.
• Per-annotation charts in the chart strip — the same charts are selectable in the Charts menu above the spans, traces, and sessions tables, alongside the overall annotation charts.
• No selection cap — the chart strip above the project tables and above the experiments table no longer limits you to three charts. Pick as many as you want to read.
• Charts load when you reach them — chart panels render a skeleton until they scroll into view, and a chart whose panel is hidden freezes its query inputs instead of refetching. A page full of annotation charts no longer fires every query at once.

​

Conversation-Grounded Hallucination Evaluator

• input and output only — input is the full record the assistant had available (its last message is the turn being answered) and output is the response being judged. There is no longer a separate context field.
• Catches fabricated work — invented specifics, tool results a tool never returned, findings from material that was never read, and actions reported as already done.
• Absence of evidence counts — a confident, fluent response that asserts situation-specific facts absent from the input is hallucinated. Ordinary general knowledge is exempt.
• As a built-in Phoenix evaluator it is promoted in the dataset evaluator gallery, and the output it judges includes the span’s tool calls.

import { createHallucinationEvaluator } from "@arizeai/phoenix-evals";
import { openai } from "@ai-sdk/openai";

const hallucinationEvaluator = createHallucinationEvaluator({
  model: openai("gpt-4o"),
});

const result = await hallucinationEvaluator.evaluate({
  input: [
    "User: What's our refund window?",
    "Tool (lookup_policy): Refunds: 30 days from delivery.",
    "Assistant: 30 days from delivery.",
    "User: And for electronics?",
  ].join("\n"),
  output: "Electronics can be returned within 90 days.",
});
console.log(result.label); // "hallucinated"

​

Annotations in Span Downloads

• OpenInference semantic attributes — span annotations are attached to their own span and trace annotations once per trace (on the root span where there is one), as indexed annotations.* and trace.annotations.* attributes carrying name, annotator kind, score, label, explanation, identifier, and JSON-encoded metadata.
• Streamed straight to disk — where the browser supports it, a download writes through a save file picker as pages arrive instead of buffering the whole export in memory.
• Parallel fetches — independent ID batches are fetched with bounded concurrency while cursor pagination within a batch stays ordered, so large selections finish substantially faster.

​

Pinned Note Bar in Span Details

• Enter adds the note, Shift+Enter starts a new line, and the field grows to six lines before it scrolls.
• Escape closes an empty bar; with a draft in it, Escape just blurs so nothing is lost.
• Stays open across spans — the bar is remembered as a preference, so a review session keeps its note field until you close it. A failed submission puts your draft back.

​

Also in This Release

• The time range follows you into a project — opening a project from the projects list carries the list’s time range along, instead of resetting to the default window (arize-phoenix 19.16.0+).
• Collapsed cards say what they hold — LLM messages, invocation parameters, LLM input, and playground and prompt chat templates show a one-line excerpt of their body in the header while collapsed (arize-phoenix 19.16.0+).
• Exception stack traces are readable — an exception span event renders its exception.stacktrace as a dedicated, expandable Stack trace card with a copy button, with the remaining attributes below it (arize-phoenix 19.16.0+).
• Tool counts in LLM span card headers — the input card subtitle shows how many tools the model had available, and the output card shows how many tool calls it made. Every span card also gets a copy button in its top-right corner (arize-phoenix 19.11.0+).
• Reasoning survives OTel conversion — reasoning parts in OTel GenAI gen_ai messages are now flattened to OpenInference message contents with type reasoning instead of being dropped (arize-phoenix 19.14.0+).
• Stricter span filter validation — malformed filter conditions, unsupported syntax, and excessively nested expressions are rejected with a clear syntax error rather than failing unpredictably (arize-phoenix 19.11.1+).
• Provider-agnostic model cost entries — creating a model in Settings → Models accepts an empty provider, and a name collision now reports the actual conflict (arize-phoenix 19.11.0+).
• Refreshed built-in token prices so cost tracking stays accurate for the current model lineup (arize-phoenix 19.11.1+).
• px setup fails when verification fails — a run that tried to confirm traces and never saw one exits 6 (NOT_VERIFIED) instead of 0, so px setup && npm run dev and set -e bootstrap scripts stop on a broken instrumentation. Choosing to verify later still exits 0.
• px suggests upgrading — an unknown command now compares your installed version against the latest published one and points at px self update when you’re behind, or px --help when you’re current (@arizeai/phoenix-cli 1.14.0+).
• px auth status no longer errors when a profile holds stale OAuth credentials but the server has since allowed anonymous access (@arizeai/phoenix-cli 1.13.1+).