feat: migrate Anthropic provider to native structured outputs API by cpsievert · Pull Request #263 · posit-dev/chatlas

cpsievert · 2026-01-06T15:53:03Z

Summary

Structured data extraction with ChatAnthropic() now uses Anthropic's native structured outputs API for Claude 4.5+ models. This means:

Streaming works with data_model — previously disabled with a warning, now fully supported
More reliable output — Anthropic's constrained decoding guarantees valid JSON matching your schema
Cleaner responses — JSON comes back as text content instead of synthetic tool calls

Older models automatically fall back to the existing tool-based approach, so this is a non-breaking change. A new structured_output_mode parameter on ChatAnthropic() and ChatBedrockAnthropic() lets you override the auto-detection — set "native" to force the output_config API or "tool" to force the legacy approach.

Also includes

Tests: sync and async streaming with data_model for Anthropic
Re-recorded VCR cassettes for data extraction tests

Test plan

test_data_extraction — verifies structured extraction via GA API
test_anthropic_nested_data_model_extraction — nested Pydantic models (regression test for Structured Data Not Working for ChatAnthropic() #100)
test_stream_with_data_model — sync streaming with data_model
test_stream_async_with_data_model — async streaming with data_model
Type checking passes
Full test suite passes

Use the new output_format parameter and beta client for models that support structured outputs (claude-sonnet-4-5, claude-opus-4-1, claude-opus-4-5, claude-haiku-4-5). This enables streaming with data_model for these models. Older models fall back to the previous tool-based approach. Documentation: https://platform.claude.com/docs/en/build-with-claude/structured-outputs 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Rename use_beta_structured_output to _supports_structured_outputs - Inline STRUCTURED_OUTPUT_MODELS into the function - Remove streaming fallback/warning for older models (let it fail) - Make create_data_model_tool a staticmethod - Revert unnecessary kwargs->api_kwargs rename in _token_count_args - Remove redundant comments about Beta API type signatures 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Move STRUCTURED_OUTPUTS_BETA and supports_structured_outputs after TYPE_CHECKING - Make supports_structured_outputs public (remove leading underscore) - Consolidate data_model handling into single block - Add TODO for removing legacy tool-based approach 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

…d-data # Conflicts: # CHANGELOG.md # chatlas/_provider_anthropic.py

- Replace explicit model allowlist with regex pattern matching (claude-*-4-[5-9]) to automatically support future Claude 4.5+ models - Switch from beta endpoint (client.beta.messages.create) to regular messages.create since structured outputs are now GA - Rename output_format to output_config.format per the GA API surface - Remove STRUCTURED_OUTPUTS_BETA constant

…f noise

- Remove type: ignore comments — output_config is in SubmitInputArgs - Remove cast(list, ...) by appending to tool_schemas before dict construction - Add OutputConfigParam import for proper type annotation

Add sync and async streaming tests with data_model for ChatAnthropic, mirroring the existing OpenAI tests. Re-record existing data extraction cassettes to use the GA output_config parameter.

…traction Models that don't support native structured outputs (pre-4.5) were silently attempting to stream with the tool-based approach after the guard was removed during the output_config refactor.

Widen supports_structured_outputs() regex to match all Claude 4 models (including 4.0, 4.1), not just 4.5+. Add "cookie" to VCR filter_headers and remove leaked _cfuvid cookies from cassette.

…ckAnthropic Lets users override auto-detection of native structured outputs support with "native" (force output_config API) or "tool" (force legacy approach). Defaults to "auto" which uses the model regex heuristic.

cpsievert changed the base branch from feat/stream-data-model to main January 6, 2026 15:58

cpsievert force-pushed the feat/stream-structured-data branch 5 times, most recently from 1f65317 to 677a184 Compare January 6, 2026 16:31

cpsievert mentioned this pull request Jan 6, 2026

Stream structured data #238

Closed

cpsievert force-pushed the feat/stream-structured-data branch from 677a184 to b268fe4 Compare January 6, 2026 16:35

cpsievert commented Jan 6, 2026

View reviewed changes