fix: serialize missing init params in to_dict (transformers, huggingface_api, ragas)

[pcbeingused333]

Related Issues

• No issue; found while auditing to_dict against __init__ across the integrations.

Proposed Changes:

Three components accept an __init__ parameter, store it, and use it at run time, but leave it out of to_dict. In each case the value silently falls back to its default after a to_dict / from_dict round trip, and nothing in the serialized dict shows that the setting was ever there. from_dict needed no change in any of them — default_from_dict passes init_parameters straight through to __init__.

TransformersZeroShotTextRouter.multi_label — passed to the Hugging Face pipeline in run:

self.pipeline([text], candidate_labels=self.labels, multi_label=self.multi_label)

The flag decides whether label scores are normalized to sum to 1 across labels, or whether each label is scored independently against its own contradiction score. The router picks its output branch from those scores, so the same text can route to a different branch after a pipeline is saved and reloaded. model and device are absent from to_dict too, but correctly so: _resolve_hf_pipeline_kwargs folds them into huggingface_pipeline_kwargs, which is serialized. multi_label is the only one that goes nowhere. The sibling component in this same integration, TransformersZeroShotDocumentClassifier, has the same shape and already serializes it — which is what suggests an oversight rather than a decision.

HuggingFaceTEIRanker.raw_scores — goes directly into the request payload, on both the sync and async paths:

payload: dict[str, Any] = {"query": query, "texts": texts, "raw_scores": self.raw_scores}

After a reload the ranker silently stops asking the TEI endpoint for raw relevance scores: the request body changes with nothing to show why. Everything else __init__ takes (url, top_k, timeout, token, max_retries, retry_status_codes) is already serialized; raw_scores is the only one missing.

RagasEvaluator.concurrency_limit — sizes the semaphore bounding how many metrics are evaluated at once in run_async:

sem = Semaphore(max(1, self.concurrency_limit))

to_dict serialized only ragas_metrics, so an evaluator built with concurrency_limit=16 comes back at the default of 4. Every one of those concurrent evaluations is an LLM call, so the value decides whether an evaluation run finishes in a quarter of the time or hits the judge model's rate limit.

How did you test it?

One new round-trip regression test per component, each asserting the value survives to_dict → from_dict:

Component	Test	Suite
TransformersZeroShotTextRouter	test_multi_label_survives_a_serialization_round_trip	hatch run test:pytest tests/test_zero_shot_text_router.py — 9 passed
HuggingFaceTEIRanker	test_raw_scores_survives_a_serialization_round_trip	hatch run test:pytest tests/test_ranker.py — 16 passed
RagasEvaluator	test_concurrency_limit_survives_a_serialization_round_trip	hatch run test:pytest tests/test_evaluator.py — 27 passed, 8 skipped (integration tests needing an API key)

hatch run test:types and hatch run fmt-check are clean in all three integrations.

I checked each regression test is not vacuous — with the production file reverted to main, each fails for the reason the bug describes rather than on an incidental assertion mismatch:

• router: assert False is True on the restored component's multi_label
• ranker: assert False is True on the restored component's raw_scores
• evaluator: assert 4 == 16 — the round trip falling back to the default

test_to_dict in the transformers and ragas suites compares the whole dict, so both needed the new key added in the same commit. The huggingface_api test_to_dict asserts key by key, which is also why that gap survived: a full-dict assertion would have flagged the missing key when raw_scores was added.

Notes for the reviewer

This supersedes #3809 (raw_scores) and #3810 (concurrency_limit), which I have closed in favour of this one. Those PRs said I would keep the fixes separate because each integration ships on its own; I have grouped them here instead, since it is one class of bug found by one sweep and it reads better reviewed as a whole than as three near-identical PRs. Releases are unaffected — CI_pypi_release.yml generates each changelog with git-cliff --include-path "<integration>/**/*", so the entry is partitioned by file path, not by the commit scope in the title.

The three commits are kept separate and per-integration, so the change can still be split back apart if you would rather take them one at a time.

concurrency_limit is documented as only affecting run_async, so that round trip is silent on the sync path — part of why it is easy to miss.

I used an AI assistant while writing this change. I have reviewed it, reproduced the behaviour, and run the tests.

Checklist

• I have read the contributors guidelines and the code of conduct
• I have added unit tests and updated the docstrings.
• I've used one of the conventional commit types for my PR title.
• I have documented my code.
• I have run pre-commit hooks and fixed any issue.

[github-actions]

Heads-up for maintainers

This PR is from a fork and touches integrations whose integration tests require API keys.
Those tests are skipped in CI because fork PRs don't have access to repo secrets for security reasons.

Affected integrations:

• huggingface_api
• ragas
• transformers

Please run the integration tests locally (hatch run test:integration inside each folder) before approving.

[github-actions]

Coverage report (transformers)

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
integrations/transformers/src/haystack_integrations/components/routers/transformers
zero_shot_text_router.py
Project Total

This report was generated by python-coverage-comment-action

[github-actions]

Coverage report (huggingface_api)

This PR does not seem to contain any modification to coverable code.

[github-actions]

Coverage report (ragas)

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
integrations/ragas/src/haystack_integrations/components/evaluators/ragas
evaluator.py
Project Total

This report was generated by python-coverage-comment-action