--- title: vector_store | CodeWeaver Docs description: API reference for codeweaver.providers.types.vector_store url: "https://docs.knitli.com/api/providers/types/vector_store" type: static generatedAt: "2026-04-17T17:21:09.399Z" --- # vector_store [Open in ChatGPT](https://chatgpt.com/?q=Read%20https%3A%2F%2Fdocs.knitli.com%2Fcodeweaver%2Fapi%2Fproviders%2Ftypes%2Fvector_store%2F.%20I%20want%20to%20ask%20questions%20about%20it.)[Open in Claude](https://claude.ai/new?q=Read%20https%3A%2F%2Fdocs.knitli.com%2Fcodeweaver%2Fapi%2Fproviders%2Ftypes%2Fvector_store%2F.%20I%20want%20to%20ask%20questions%20about%20it.)[View in Markdown](/codeweaver/api/providers/types/vector_store.md) [Share on LinkedIn](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fdocs.knitli.com%2Fcodeweaver%2Fapi%2Fproviders%2Ftypes%2Fvector_store%2F)[Share on X](https://x.com/intent/tweet?url=https%3A%2F%2Fdocs.knitli.com%2Fcodeweaver%2Fapi%2Fproviders%2Ftypes%2Fvector_store%2F&text=vector_store)[Share on Threads](https://threads.net/intent/post?url=https%3A%2F%2Fdocs.knitli.com%2Fcodeweaver%2Fapi%2Fproviders%2Ftypes%2Fvector_store%2F&text=vector_store)[Share on Bluesky](https://bsky.app/intent/compose?text=vector_store%20https%3A%2F%2Fdocs.knitli.com%2Fcodeweaver%2Fapi%2Fproviders%2Ftypes%2Fvector_store%2F)[Share on Facebook](https://www.facebook.com/sharer/sharer.php?u=https%3A%2F%2Fdocs.knitli.com%2Fcodeweaver%2Fapi%2Fproviders%2Ftypes%2Fvector_store%2F)[Share on Reddit](https://reddit.com/submit?url=https%3A%2F%2Fdocs.knitli.com%2Fcodeweaver%2Fapi%2Fproviders%2Ftypes%2Fvector_store%2F&title=vector_store)[Share on Hacker News](https://news.ycombinator.com/submitlink?u=https%3A%2F%2Fdocs.knitli.com%2Fcodeweaver%2Fapi%2Fproviders%2Ftypes%2Fvector_store%2F&t=vector_store)[Share on Email](mailto:?subject=vector_store&body=https%3A%2F%2Fdocs.knitli.com%2Fcodeweaver%2Fapi%2Fproviders%2Ftypes%2Fvector_store%2F)[Share on WhatsApp](https://wa.me/?text=vector_store%20https%3A%2F%2Fdocs.knitli.com%2Fcodeweaver%2Fapi%2Fproviders%2Ftypes%2Fvector_store%2F)[Share on Telegram](https://t.me/share/url?url=https%3A%2F%2Fdocs.knitli.com%2Fcodeweaver%2Fapi%2Fproviders%2Ftypes%2Fvector_store%2F&text=vector_store) # `codeweaver.providers.types.vector_store` [Section titled “codeweaver.providers.types.vector_store”](#codeweaverproviderstypesvector_store) Collection metadata models for vector stores. ## Class: `CollectionMetadata` [Section titled “Class: CollectionMetadata”](#class-collectionmetadata) Metadata stored with collections for validation and compatibility checks. Version History: - v1.2.0: Initial schema with dense_model, sparse_model - v1.3.0: Added dense_model_family and query_model for asymmetric embedding support - v1.4.0: Added configuration tracking and transformation tracking fields - v1.5.0: Added collection policy system for controlling configuration changes Migration from v1.2.x to v1.3.0: Collections created with v1.2.x are fully compatible with v1.3.0. The new fields (dense_model_family, query_model) default to None, indicating single-model mode: - dense_model_family=None: No model family tracking (backward compatible) - query_model=None: Symmetric mode (query uses same model as dense_model) Existing collections can be loaded, modified, and saved without requiring migration. The pydantic validators handle missing fields gracefully via Field defaults. Migration from v1.3.0 to v1.4.0: Collections created with v1.3.0 are fully compatible with v1.4.0. The new fields default to None or empty lists, maintaining backward compatibility: - profile_name, profile_version, config_hash, config_timestamp: Optional tracking - quantization_type, quantization_rescore, original_dimension: Optional transformations - transformations: Empty list by default Migration from v1.4.0 to v1.5.0: Collections created with v1.4.0 are fully compatible with v1.5.0. The new policy field defaults to FAMILY_AWARE, maintaining existing behavior: - policy: Defaults to CollectionPolicy.FAMILY_AWARE for backward compatibility - Existing collections will use FAMILY_AWARE policy unless explicitly changed - No breaking changes to existing validation logic Asymmetric Embedding Support (v1.3.0+): When dense_model_family is set, the collection supports cross-model querying within the same family (e.g., Voyage-4 family allows voyage-4-large for embedding and voyage-4-nano for queries). This enables: - Local query models (zero cost, instant latency) - API embedding models (higher quality) - 3-point retrieval improvement (per Voyage AI) Transformation Tracking (v1.4.0+): Tracks quantization and dimension reduction transformations applied to collections, enabling migration auditing and accuracy impact analysis. ### Method: `from_collection` [Section titled “Method: from_collection”](#method-from_collection) ** ``` from_collection() ``` Create CollectionMetadata from a collection dictionary. ### Method: `to_collection` [Section titled “Method: to_collection”](#method-to_collection) ** ``` to_collection() ``` Convert to a dictionary that is the argument for collection creation. ### Method: `validate_compatibility` [Section titled “Method: validate_compatibility”](#method-validate_compatibility) ** ``` validate_compatibility() ``` Validate collection metadata against current provider configuration. Performs family-aware validation when dense_model_family is present, allowing asymmetric embedding configurations where query models can differ from embed models as long as they belong to the same model family. Args: other: Other collection metadata to compare against (typically from existing collection) Raises: ModelSwitchError: If embedding models don’t match and family validation fails ConfigurationError: If models are incompatible within their family Warnings: Logs warning if provider has changed (suggests reindexing) ### Method: `validate_config_change` [Section titled “Method: validate_config_change”](#method-validate_config_change) ** ``` validate_config_change() ``` Validate configuration change against collection policy. Validates whether proposed configuration changes are allowed based on the collection’s policy setting. This method is called before applying any configuration changes to ensure they won’t break the collection. Args: new_dense_model: Proposed new dense embedding model new_query_model: Proposed new query model (for asymmetric embedding) new_sparse_model: Proposed new sparse embedding model new_provider: Proposed new provider Raises: ConfigurationLockError: If the proposed change violates the collection policy - STRICT: Any model or provider change - FAMILY_AWARE: Model change outside the same family - FLEXIBLE: Only warns, doesn’t raise - UNLOCKED: Never raises Examples: >>> metadata = CollectionMetadata( … provider=“voyage”, … project_name=“my-project”, … dense_model=“voyage-4-large”, … dense_model_family=“voyage-4”, … policy=CollectionPolicy.FAMILY_AWARE, … ) >>> # This succeeds - same family >>> metadata.validate_config_change(new_query_model=“voyage-4-nano”) >>> # This raises - different family >>> metadata.validate_config_change(new_dense_model=“voyage-3-large”) Traceback (most recent call last): … ConfigurationLockError: Model change breaks family compatibility ## Class: `CollectionPolicy` [Section titled “Class: CollectionPolicy”](#class-collectionpolicy) Collection modification policy controlling configuration changes. Defines how strictly a collection’s configuration is enforced when attempting to use different embedding models or providers with an existing collection. Values: STRICT: No model changes allowed - only the exact original model can be used. FAMILY_AWARE: Allow query model changes within the same model family (default). This is the recommended setting as it enables asymmetric embedding where documents are embedded with a high-quality model (e.g., voyage-4-large) and queries use a lighter model (e.g., voyage-4-nano) for better latency/cost. FLEXIBLE: Warn on breaking changes but don’t block them. Use with caution as this can lead to degraded search quality or errors. UNLOCKED: Allow all configuration changes without validation. Only use for testing or when you fully understand the implications. Default: FAMILY_AWARE - balances safety with flexibility for asymmetric embedding. Examples: >>> # Strict policy - no changes allowed >>> metadata = CollectionMetadata( … provider=“voyage”, project_name=“my-project”, policy=CollectionPolicy.STRICT … ) >>> # Will raise ConfigurationLockError if model changes >>> # Family-aware policy (default) - allows query model changes in family >>> metadata = CollectionMetadata( … provider=“voyage”, … project_name=“my-project”, … dense_model=“voyage-4-large”, … dense_model_family=“voyage-4”, … policy=CollectionPolicy.FAMILY_AWARE, # or omit for default … ) >>> # Can use voyage-4-nano for queries, but not voyage-3 models ## Class: `HybridVectorPayload` [Section titled “Class: HybridVectorPayload”](#class-hybridvectorpayload) Metadata payload for stored vectors. ### Method: `from_payload` [Section titled “Method: from_payload”](#method-from_payload) ** ``` from_payload() ``` Create a HybridVectorPayload from a dictionary payload. ### Method: `index_field_types` [Section titled “Method: index_field_types”](#method-index_field_types) ** ``` index_field_types() ``` Return the payload fields mapped to their datatype for Qdrant indexing. ### Method: `indexed_fields` [Section titled “Method: indexed_fields”](#method-indexed_fields) ** ``` indexed_fields() ``` Return the payload fields that are indexed by default in the Qdrant collection. ### Method: `to_payload` [Section titled “Method: to_payload”](#method-to_payload) ** ``` to_payload() ``` Convert to a dictionary payload for storage. ## Class: `PayloadFieldDict` [Section titled “Class: PayloadFieldDict”](#class-payloadfielddict) A mapping of payload field names to their corresponding Qdrant data types; not all fields are indexed, but if they were, these would be the types. ## Class: `TransformationRecord` [Section titled “Class: TransformationRecord”](#class-transformationrecord) Record of a transformation applied to collection. Tracks changes like quantization or dimension reduction applied to a collection, including timing and accuracy impact metrics. This dataclass is immutable to ensure transformation records cannot be modified after creation.