feat: export private collection config types as public aliases

[LijuanTang94]

Summary

Closes #1994

Users who want to type-annotate parameters like generative_config, reranker_config, or vector_index_config in collection.create() and collection.config.update() currently have to import private _-prefixed types. Under pyright strict (reportPrivateUsage: error) that is an error.

This PR adds public TypeAlias entries for every private type in both tables of #1994, re-exports them from weaviate.classes.config, switches the executor signatures to the public names, and regenerates the sync.pyi / async_.pyi stubs — the same shape as #1993 did for FilterReturn.

23 new public exports (26 after the follow-up below; struck rows were dropped, see below):

create side	update side
GenerativeProvider	InvertedIndexConfigUpdate
InvertedIndexConfigCreate	MultiTenancyConfigUpdate
MultiTenancyConfigCreate	NamedVectorConfigUpdate
NamedVectorConfigCreate	ObjectTTLConfigUpdate
ObjectTTLConfigCreate	NamedVectorConfigUpdate
ReferencePropertyBase	VectorConfigUpdate
ReferencePropertyMultiTarget	VectorIndexConfigDynamicUpdate
ReplicationConfigCreate	VectorIndexConfigFlatUpdate
RerankerProvider	VectorIndexConfigHFreshUpdate
ShardingConfigCreate	VectorIndexConfigHNSWUpdate
VectorConfigCreate
VectorIndexConfigCreate
VectorizerConfigCreate

Before

# pyright error: reportPrivateUsage
from weaviate.collections.classes.config import _GenerativeProvider

def my_helper(gen: Optional[_GenerativeProvider]) -> ...:
    ...

After

from weaviate.classes.config import GenerativeProvider

def my_helper(gen: Optional[GenerativeProvider]) -> ...:
    ...

Result

No private config type remains in any published stub — _GenerativeProvider, _RerankerProvider, _ReferencePropertyBase, _VectorizerConfigCreate, _NamedVectorConfigCreate, _VectorConfigCreate, _InvertedIndexConfigUpdate and _ReplicationConfigUpdate all now appear in 0 .pyi files. The signatures read:

generative_config: Optional[GenerativeProvider] = None,
references: Optional[List[ReferencePropertyBase]] = None,
inverted_index_config: Optional[InvertedIndexConfigUpdate] = None,
replication_config: Optional[ReplicationConfigUpdate] = None,

Testing

ruff check / ruff format --check   clean
pyright 1.1.399 (weaviate/)        0 errors
pytest test mock_tests             490 passed, 1 skipped

Annotating every create() argument in a standalone file under pyright strict with reportPrivateUsage: error: 12 errors using the private names, 0 using the aliases.

Credit

The update-side aliases, the three renames (GenerativeProvider, RerankerProvider, ReferencePropertyBase), the executor signature switch and the stub regeneration come from @g-despot's follow-up branch, merged into this one.

Follow-up changes

Pushed on top of the above, plus a merge of main:

• The public name is now the class, not an alias to it. An alias is transparent, so the private name stayed authoritative wherever it is actually rendered — _validate_input error text, repr/__name__, IDE hovers, and Sphinx, which showed each alias as a bare alias of _X with no fields. Flipped for 29 names; _X: TypeAlias = X shims keep every old spelling importable and identical.

• Create/update symmetry. The table above exports the create base with none of its 5 leaves, and 4 update leaves with no base. Added VectorIndexConfigUpdate and the 5 VectorIndexConfig*Create leaves.

• All 23 are now reachable from weaviate.collections.classes.config, which previously re-exported the private counterparts of 12 of them and none of the public names.

• boost uses the public BoostReturn instead of _Boost in the 14 query executors. BoostReturn already existed and was already exported from weaviate.classes.query.

• Correction to "nothing removed": 54 private names are no longer re-exported incidentally from executor module namespaces (plus 14 more from the _Boost swap). Checked all 290 importable modules — no private name became unimportable, every one keeps its defining module, and nothing in this repo imported the dropped spellings.

• The three deprecated-path names are not exported. VectorizerConfigCreate, NamedVectorConfigCreate and NamedVectorConfigUpdate annotate nothing but deprecated paths — vectorizer_config emits Dep024/Dep023 and the named-vector add_vector overload is @deprecated, and both warnings point at vector_config instead. They stay private so the signature reads as the discouragement it should:

  vectorizer_config: Optional[Union[_VectorizerConfigCreate, List[_NamedVectorConfigCreate]]]  # deprecated
  vector_config:     Optional[Union[VectorConfigCreate, List[VectorConfigCreate]]]             # use this

• Changelog entry added for 4.24.0.

Out of scope: the factory boundary — Configure.VectorIndex.hnsw() still takes quantizer: Optional[_QuantizerConfigCreate]. Every normal way of building a quantizer (inline, bound to a variable, a dict of presets, a helper with no return annotation) type-checks clean under strict, so this only bites on an explicit annotation. Worth its own PR.

Still missing: a test for the import surface. A rebind typo would type-check, pass every test and stub regen, and ship.

On the merged head: pyright 1.1.399 0 errors, pytest test mock_tests 506 passed / 1 skipped, ruff + flake8 clean, stub regeneration idempotent. 26 public exports.

🤖 Generated with Claude Code

https://claude.ai/code/session_018mG6vBbjJfEVxaH95bwnVh

[orca-security-eu]

Orca Security Scan Summary

Status	Check	Issues by priority
Passed	Infrastructure as Code	0   0   0   0	View in Orca
Passed	SAST	0   0   0   0	View in Orca
Passed	Secrets	0   0   0   0	View in Orca
Passed	Vulnerabilities	0   0   0   0	View in Orca

[weaviate-git-bot]

To avoid any confusion in the future about your contribution to Weaviate, we work with a Contributor License Agreement. If you agree, you can simply add a comment to this PR that you agree with the CLA so that we can merge.

beep boop - the Weaviate bot 👋🤖

PS:
Are you already a member of the Weaviate Forum?

[LijuanTang94]

I have read the CLA and agree to the terms.

[LijuanTang94]

Hi team — just checking in on this PR. CLA is now signed and all checks are passing. Happy to make any changes if needed!

[LijuanTang94]

Rebased onto current main (was 93 commits behind) — no conflicts, diff unchanged from the original.

Same CI situation as #2050: the Main workflow has never run here either — it's stuck at action_required pending first-time-contributor approval, so only the security scans show. Approving the run would make the real checks visible.

Locally against current main: 415 passed on main, 415 on this branch — no regression, as expected for a pure re-export. All 8 private types still exist under those names on main, and all 8 aliases resolve and land in weaviate.classes.config.__all__. ruff format --check is clean.

On the change: #1994 is still open, and this follows the same pattern as #1993 (FilterReturn) — TypeAlias entries so users running pyright in strict mode don't have to import _-prefixed types just to annotate collection.create() arguments.

@dirkkul

[Copilot]

🟡 Changes recommended

Public signatures remain private and several creation types covered by #1994 are omitted.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Adds public aliases for collection creation configuration types to avoid private imports.

Changes:

• Defines eight public TypeAlias entries.
• Re-exports them through weaviate.classes.config.

File summaries

File	Description
weaviate/collections/classes/config.py	Defines six configuration aliases.
weaviate/collections/classes/config_object_ttl.py	Defines the object TTL alias.
weaviate/collections/classes/config_vector_index.py	Defines the vector index alias.
weaviate/classes/config.py	Exposes all eight aliases publicly.

Review details

• Files reviewed: 4/4 changed files
• Comments generated: 2
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[LijuanTang94]

Thanks for the review. Both of Copilot's points are correct — I've fixed the second and scoped the PR to match on the first.

Missing aliases. Right: the first table in #1994 lists eleven create() arguments and I had only covered eight. 5167bbe adds the four that were missing:

Public alias	Private type	create() argument
ReferencePropertyCreate	_ReferencePropertyBase	references
VectorizerConfigCreate	_VectorizerConfigCreate	vectorizer_config
NamedVectorConfigCreate	_NamedVectorConfigCreate	vectorizer_config
VectorConfigCreate	_VectorConfigCreate	vector_config

To confirm the aliases actually solve the reported problem, I annotated all eleven create() arguments in a standalone file under pyright strict with reportPrivateUsage: error: 12 errors using the private names, 0 using the aliases.

Signatures still expose the private names. Also correct, and I've updated the PR description to say "Part of #1994" rather than "Closes". Re-reading #1993, it does go further than this PR: alongside the __all__ re-export it swapped _Filters → FilterReturn in every executor and regenerated all the .pyi stubs, ~60 files.

I've kept this one to the alias export because it's independently useful and easy to review, and listed the remainder in the description as follow-ups:

1. switch the collections/collections/executor.py annotations to the public names and regenerate sync.pyi / async_.pyi;
2. the second table in Private types exposed in public method signatures #1994 — the whole collection.config.update() path (_InvertedIndexConfigUpdate, _MultiTenancyConfigUpdate, _ObjectTTLConfigUpdate, _ReplicationConfigUpdate, _VectorIndexConfig*Update, _NamedVectorConfigUpdate, _VectorConfigUpdate), which is equally in scope for the issue.

Happy to fold either or both into this PR if you'd rather land it as a single change — just let me know before I touch ~60 files.

One ask: CI here is still sitting at action_required, so only the security scans have ever run on this branch. You approved the run on #2050 — could you do the same here? Locally against current main:

ruff check / ruff format --check   clean
flake8                             clean
pyright 1.1.399 (weaviate/)        0 errors
pytest test/                       434 passed on main, 434 on this branch

Unchanged test count, as expected for a pure re-export.

[g-despot]

Thanks a lot for taking care of the comments, I have created a PR on top of yours that contains three minor renames and usage of the new names in method signatures.

[Copilot]

🟡 Changes recommended

The new generative annotation types remain absent from the intended public generate namespace, and the export surface lacks regression coverage.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

Suppressed comments (3)

weaviate/collections/classes/config.py:3143

• Use “a” before the newly public class name: “VectorIndexConfigHFreshUpdate” begins with a consonant sound.

        """Create an `VectorIndexConfigHFreshUpdate` object to update the configuration of the HFresh vector index.

weaviate/collections/classes/config.py:3103

• Use “a” before the newly public class name: “VectorIndexConfigFlatUpdate” begins with a consonant sound.

        """Create an `VectorIndexConfigFlatUpdate` object to update the configuration of the FLAT vector index.

weaviate/collections/classes/config.py:3123

• Use “a” before the newly public class name: “VectorIndexConfigDynamicUpdate” begins with a consonant sound.

        """Create an `VectorIndexConfigDynamicUpdate` object to update the configuration of the Dynamic vector index.

• Files reviewed: 62/62 changed files
• Comments generated: 3
• Review effort level: Balanced