arize-phoenix-v14.0.0
Breaking Changes📦 arize-phoenixView on GitHub →
⚠ 4 breaking✨ 19 features🐛 3 fixes⚡ 2 deprecations🔧 6 symbols
Summary
Version 14.0.0 introduces major architectural changes, including a subcommand-first CLI structure and the removal of legacy evaluation and client plumbing. Key features include enhanced agent tooling, PostgreSQL read replica support, and ephemeral experiment handling.
⚠️ Breaking Changes
- The /v1/evaluations endpoint and associated Evaluations plumbing have been removed from the server.
- Evals 1.0 has been deprecated and the legacy experiments module has been removed.
- CLI flags now follow the subcommand instead of preceding it. Example: Change 'phoenix --dev serve' to 'phoenix serve --dev'.
- The legacy client has been removed.
Migration Steps
- Review MIGRATION.md for detailed upgrade instructions.
- Update CLI commands to place flags after the subcommand (e.g., use 'phoenix serve --dev' instead of 'phoenix --dev serve').
- If using deprecated Evals 1.0 functionality, migrate to the new system.
- Remove any usage of the removed /v1/evaluations endpoint.
- Remove any usage of the removed legacy client.
✨ New Features
- Added support for [@defer] to Relay and ProjectPageHeader.
- Added read replica routing support for PostgreSQL.
- Added Strands Agents integration docs and onboarding snippet.
- Agent: Added ask_user elicitation tool with carousel UI.
- Agent: Added backend MCP docs tool via Mintlify integration.
- Agent: Chat streaming is maintained when the panel closes.
- Agent: Implemented PXI FAB, panel sizing, and stop button.
- Agent: Traced PXI chat requests locally and remotely.
- Implemented background experiment runner.
- CLI: Restructured CLI into subcommand-first architecture with serve/db commands.
- Experiments: Defined ephemeral experiments and added ExperimentSweeper daemon.
- Experiments: Implemented structured error table with sanitized messages.
- GraphQL: Require explicit 'first' argument for forward pagination.
- Included span and experiment run in error subscription payloads.
- Onboarding: Added new provider integrations and environment variable support.
- Playground: Added experiment record toggle with ephemeral tracking.
- Implemented polling for the experiments table when jobs are running.
- Using non-streaming for LLM evaluator calls.
- Using raw vendor response as output.value for non-streaming playground.
🐛 Bug Fixes
- Experiments: Filtered errors by level, used LATERAL join, and renamed to ExperimentLog.
- Experiments: Fixed lastError query failing on SQLite.
- Persisted error records for all terminal experiment outcomes.
Affected Symbols
⚡ Deprecations
- Evals 1.0 is deprecated.
- Legacy experiments module is removed.