fix: make Document.id deterministic regardless of meta key order

[Aarkin7]

Related Issues

• fixes fix: make Document.id deterministic regardless of meta key order #11445

Proposed Changes:

Fixes a bug where Document.id depended on the insertion order of keys in meta. The hash was built from dict's repr, which reflects insertion order, so two Documents with the same content and the same metadata could end up with different IDs, silently breaking DuplicatePolicy.SKIP/FAIL and any cache keyed on the document ID.

Serializing meta with json.dumps(..., sort_keys=True) before hashing makes the ID order-independent. Empty-meta IDs are
preserved by keeping the legacy "{}" string, so only documents with non-empty meta get new IDs.

How did you test it?

• Added two regression tests in test/dataclasses/test_document.py covering flat and nested meta key orderings.
• Updated two stale hardcoded IDs in existing tests (test_init_with_parameters, test_init_with_legacy_field) since the hash changes for non-empty meta.
• hatch run test:unit → 5135 passed, 0 failed.
• hatch run test:types → clean.
• hatch run fmt → clean.

Notes for the reviewer

• Breaking-ish: any auto-generated Document.id for documents with non-empty meta changes. The release note's upgrade section calls this out and tells users how to handle it (re-ingest, or pass id explicitly). Empty-meta IDs are intentionally preserved to minimize churn.
• Scope was kept tight: the bug report mentions embedding and sparse_embedding hashing as related concerns. SparseEmbedding.to_dict() is order-stable today via dataclass field order, and float-embedding nondeterminism isn't fixable at this layer, so this PR sticks to the actual reported root cause.
• default=str in json.dumps is a safety net for the (unsupported but tolerated) case of non-JSON-serializable values in meta, matching the prior behavior of not crashing on such inputs.

Checklist

• I have read the contributors guidelines and the code of conduct.
• I have updated the related issue with new insights and changes.
• I have added unit tests and updated the docstrings.
• I've used one of the conventional commit types for my PR title: fix:, feat:, build:, chore:, ci:, docs:, style:, refactor:, perf:, test: and added ! in case the PR includes breaking changes.
• I have documented my code.
• I have added a release note file, following the contributors guidelines.
• I have run pre-commit hooks and fixed any issue.

[vercel]

@Aarkin7 is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
haystack/dataclasses
document.py
Project Total

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

[biswajeetdev]

The core fix is right — str(dict) includes Python insertion order in its output, so two Document objects with identical content and logically identical meta dicts could get different IDs just because the dicts were constructed differently. json.dumps(..., sort_keys=True) produces a canonical string that is order-independent.

One thing worth calling out in the release note: the default=str argument also affects documents with non-JSON-serialisable meta values (e.g. datetime objects, custom classes). Previously str({"date": datetime(2024,1,1)}) produced "datetime.datetime(2024, 1, 1, 0, 0)" in the hash; now json.dumps(..., default=str) produces "2024-01-01 00:00:00". These are different strings, so any documents with non-serialisable meta fields will also get new IDs — not just documents where key order varied. The release note currently says only non-empty-meta documents are affected, which is true but understates the impact for users with datetime or custom-type meta.

The if self.meta else "{}" guard correctly matches str({}) == "{}" for empty-dict meta, so the "empty meta is unaffected" claim holds.

[Aarkin7]

Thanks for catching this. You're right that the impact wording was too narrow. Updated the upgrade note to spell out both cases explicitly:

• documents with non-empty meta (repr -> JSON serialization change)
• documents whose meta contains non-JSON-serializable values like datetime or custom classes (now serialized via str(...) instead of repr(...) thanks to default=str)