weaviate-client

Provides dual WeaviateClient (sync) and WeaviateAsyncClient (async) classes that abstract HTTP connection management to a Weaviate vector database instance. Both inherit from _WeaviateClientExecutor base class implementing shared core functionality, with connection parameters (host, port, protocol) passed via ConnectionParams objects. Supports embedded Weaviate instances via EmbeddedOptions, custom headers, authentication credentials, and configurable timeouts through AdditionalConfig. Initialization can skip server health checks via skip_init_checks flag for faster startup in trusted environments.

Exposes client.collections namespace for CRUD operations on Weaviate schema classes (collections). Allows creating collections with dynamic property definitions, vectorization settings (module selection), and indexing strategies without manual schema validation. Collections are created via fluent API accepting property objects with data types, vectorization hints, and indexing parameters. Supports retrieving existing collections, updating collection settings, and deleting collections with cascade options. Schema validation is performed server-side with detailed error messages returned to client.

Exposes client.cluster namespace for inspecting Weaviate cluster topology and node health. Provides methods to list cluster nodes, retrieve node status (healthy/unhealthy), and inspect node metadata (shard count, vector count, memory usage). Node status is retrieved from Weaviate server and reflects current cluster state. No cluster modification operations are supported via client — cluster topology is managed via Weaviate server configuration.

Supports embedded Weaviate instances via EmbeddedOptions, allowing developers to run Weaviate in-process without separate server. Embedded instance is started automatically on client initialization and stopped on client close. Supports configurable persistence (in-memory or disk-backed), port binding, and data directory. Embedded Weaviate is fully functional — supports all client operations (collections, queries, batch import) with same API as remote instances. Useful for local development, testing, and prototyping without Docker/Kubernetes overhead.

Supports configurable vectorization modules (text2vec-openai, text2vec-huggingface, text2vec-cohere, etc.) at collection level, enabling automatic embedding generation for text properties. Vectorization module is selected at collection creation and applied to specified properties. Client does not perform embedding generation — Weaviate server handles vectorization using configured module and provider credentials. Supports per-property vectorization configuration (which properties trigger embedding, which skip). Vectorization is transparent to client — objects are inserted with text, embeddings are generated server-side.

Supports reference properties that create relationships between objects in different collections, enabling graph-like queries. References are defined at collection creation with target collection specification. Objects are inserted with reference values (target object IDs). Queries can traverse references via client.collections[name].query.near_vector().with_references() to include related objects in results. References are server-side relationships — no client-side graph construction. Supports bidirectional reference queries.

Implements comprehensive error handling via custom exception classes (WeaviateConnectionError, WeaviateInvalidInputError, WeaviateAuthenticationError, etc.) that map Weaviate server errors to Python exceptions. Error messages include server-side error details, HTTP status codes, and suggested remediation. Supports error recovery patterns (retry logic, connection pooling) at client level. Error handling is transparent — client code catches specific exceptions rather than parsing HTTP responses.

Implements vector search via client.collections[name].query.near_vector() method, accepting a query vector and returning ranked results based on distance metric (cosine, L2, dot product, hamming). Search results include object data, distance scores, and optional metadata. Supports limiting result count, offset pagination, and result sorting by distance or other properties. Distance metric is configured at collection creation time and applied consistently across all queries. Results are returned as typed objects matching collection schema.

Provides client.collections[name].query.hybrid() method combining BM25 keyword search with vector similarity search, returning fused results ranked by configurable alpha parameter (0=pure BM25, 1=pure vector, 0.5=equal weight). Internally executes both keyword and vector queries against Weaviate server, then merges result sets using reciprocal rank fusion or other fusion algorithms. Supports filtering, limiting, and sorting on fused results. BM25 parameters (k1, b) are configured server-side and applied consistently.

Implements client.collections[name].query.near_vector().with_generate() method that chains vector search with LLM-based generation, using retrieved objects as context. Supports multiple generation modes: single prompt applied to all results, per-result prompts, and grouped generation. LLM provider (OpenAI, Cohere, Hugging Face, etc.) is configured at collection level via generative module. Generated text is returned alongside search results without additional API calls from client. Supports prompt templating with result field substitution.

Exposes client.batch namespace for optimized bulk data import via batch_objects() method, supporting multiple batching strategies: fixed-size batches, dynamic batching with timeout, and streaming mode. Internally queues objects and sends them in optimized HTTP requests to Weaviate server, with configurable batch size (default 100) and timeout (default 60s). Supports error handling modes: fail-fast, continue-on-error with error collection, and retry logic for transient failures. Returns batch results with per-object status (success/failure) and error details.

Provides where() method on query builders (near_vector, hybrid, etc.) accepting filter objects that compile to Weaviate WHERE clause JSON. Supports complex boolean logic (AND, OR, NOT) with nested conditions, comparison operators (==, !=, >, <, >=, <=, ~), and special operators (contains, like, in). Filters are applied server-side before vector search, reducing result set before ranking. Supports filtering on any collection property (text, number, date, reference). Filter compilation is transparent — client code uses Python objects, not raw JSON.

Enables multi-tenant data isolation via collection-level tenant configuration, where each tenant's data is logically partitioned within a single collection. Tenants are created via client.collections[name].tenants.create() and data is inserted with tenant context via batch_objects(tenant='tenant_id'). Queries automatically scope to specified tenant via near_vector(tenant='tenant_id'). Tenant isolation is enforced server-side — no cross-tenant data leakage. Supports tenant deletion with cascade options.

Exposes client.roles and client.users namespaces for managing Weaviate RBAC system. Roles are created with specific permissions (read, write, delete, manage_rbac) on collections or cluster-wide. Users are assigned to roles, inheriting permissions. Permissions are enforced server-side on all API calls. Supports role creation, permission assignment, user-role binding, and role deletion. RBAC is optional — disabled by default, enabled via Weaviate server configuration.

Provides client.backup namespace for creating and restoring Weaviate backups across multiple storage backends. Supports filesystem (local disk), S3 (AWS), GCS (Google Cloud), and Azure Blob Storage as backup destinations. Backup creation is asynchronous — client initiates backup and polls for completion status. Restore operations restore entire Weaviate instance or specific collections from backup. Backup metadata (timestamp, size, collections included) is returned with backup status.