pypi / weaviate-client
weaviate-client
What changed in weaviate-client 4, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-28.
1 breaking1 security2 deprecations18 new APIs and options1 behavior and defaults15 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli initweaviate-client 438 changes
Breaking changes 1
Multi-target searches reject mismatched vector names.
Security 1
gRPC channels now trust certifi's CA bundle.
Deprecations 2
graphql_raw_query is deprecated.
The replication argument async_enabled is deprecated because it was removed server-side.
New APIs and options 18
The client supports dropping a vector index through alter schema.
backup.create accepts include_roles as input.
backup.create accepts include_users as input.
Generative-meta support is added.
A pathseer filter strategy is available for the HNSW vector index.
Private collection config types are exported as public aliases for use by developers.
The async client supports a grpc-web transport for Pyodide/WASM.
Read-only backup permissions are supported.
Exposes frequency_penalty on the Cohere runtime config.
Exposes anthropic base_url on the collection config.
Exposes nvidia top_p on the collection config.
Exposes api_version on the Azure OpenAI generative configuration.
Adds location to the Google Vertex generative configuration.
Adds generative DigitalOcean support.
Adds location to the Google Vertex runtime configuration.
Exposes n on the OpenAI runtime configuration.
Adds support for RQ4 centered quantization.
All connection helpers now accept string API keys.
Behavior and default changes 1
The method checks shard indexing status per node while waiting for vector indexing.
Fixes and other changes 15
Cached binaries are isolated for custom URLs.
Exhausted object iterators no longer fetch again.
Replication now parses structured error messages.
The method no longer loses multivector configuration when converting collection config to a dictionary.
Unspecified offsets are omitted from date-property TTL updates.
The HFresh multivector config is now read from the server.
Object TTL updates are written under objectTtlConfig.
The method now reads the boolean response body.
ShardTypes now lists every shard status.
Token refresh lifecycle, backoff, and error-taxonomy follow-ups were fixed.
AWS stop_sequences are now sent through to the gRPC request.
Requires both counts to remain unchanged before an idle flush.
Fixes rate-limit scheduling for partial batches.
Fixes KeyError 'pq' when updating the quantizer on an HFresh vector index.
Fixes reusable reference filter builders.