v0.5.0 — multi-track enhancements

v0.5.0 — multi-track enhancements

Five independent feature tracks, each with its own tests and a backward-compatible contract. +53 new tests, 994 passing locally, 0 regressions.

Track A — International LLM providers

Adds Groq, Mistral, Together AI, Cohere, and Fireworks via INTERNATIONAL_PROVIDER_PRESETS (mirroring the CHINA_PROVIDER_PRESETS pattern). Resolved as get_backend("intl:groq" | "intl:mistral" | "intl:together" | "intl:cohere" | "intl:fireworks"). All OpenAI Chat-Completions compatible, shared OpenAICompatBackend under the hood. Multimodal variants auto-picked with multimodal=True. Total registered backends: 23 (was 18).

Track B — New document types

• W2 (US wage and tax statement) with W2LineItem + templates/w2.md
• IDCard (driver's license / passport / state ID) + templates/id_card.md
• MedicalClaim (CMS-1500 / UB-04) with MedicalClaimLineItem + templates/medical_claim.md

All three ship with Markdown templates (frontmatter + field descriptions + "Common mistakes" section) and register in SCHEMA_REGISTRY so Pipeline(schema="W2"), extract_one(..., schema="W2"), and template auto-discovery all work out of the box.

Track C — Eval harness per-field report + CI regression gate

run_dataset() now additionally returns per_field_accuracy (aggregated across all cases × all strategies). New check_regression(results, thresholds) returns failing threshold descriptions; verify_against_thresholds() raises EvalRegressionError for CI. New format_report(results) is a plain-text pretty-printer. Threshold grammar: per_strategy.<name>.<metric> and per_field_accuracy.<field>.<metric>.

A new CI workflow .github/workflows/eval-regression.yml wires the gate end-to-end with deliberately-permissive 0.0 thresholds for v0.5 (tighten in v0.6 once real backends have a baseline).

Track D — Async pipeline

AsyncPipeline wraps the existing Pipeline so callers can await apipe.arun(doc) and await apipe.arun_many(docs). New Backend.acomplete() method defaults to asyncio.to_thread(self.complete, req) so all shipped backends transparently support await with zero risk to existing code. process_batch_async() in batch.py mirrors the sync helper with a bounded concurrency= semaphore and an on_progress callback. BatchItemResult gains .success and .latency_s fields, kept consistent with existing .ok / .elapsed_seconds via __post_init__.

A new example notebook examples/notebooks/04_async_pipeline.ipynb walks through all the async surface in 17 cells — all runnable with MockBackend, no API key needed.

Track E — Registry helpers

list_backends() enumerates every registered backend as {name, family, auth} dicts. Gated backends (slowmock, nanonets) surface only when their IDP_ENABLE_* env var is set. Backend.__repr__ defaults to <ClassName name=...> for nicer debug logs.

Backward compatibility

Every change is additive. No public API renames or removals. Sync process_batch callers see no behavior change from the BatchItemResult field additions. All public re-exports are additive (3 in idp.llm, 3 in idp.eval, 1 in idp.pipeline).

Test coverage

Test file	Tests
tests/test_international_providers.py	6
tests/test_new_schemas.py	11
tests/test_eval_regression.py	12
tests/test_async_pipeline.py	14
tests/test_backend_registry_helpers.py	10
New tests total	53

Full suite: 994 passed, 2 skipped (pre-existing test_enhance_benchmark.py needs pytest-benchmark plugin), 0 regressions. python -c "import idp" still works.

Install

pip install --upgrade py-idp

Or pin: pip install py-idp==0.5.0

Acknowledgments

This release was assembled with parallel sub-agent work streams (4 feature tracks + 1 integration track + 2 follow-on tracks) by the maintainer's local agent harness.