Last 7 days
2
Features: 0
Changes: 0
Fixes: 2
Deprecations: 0
Weaviate
Open-source vector database for AI-native application retrieval.
Latest Weaviate changelog updates, official release notes, breaking changes, security patches, pricing changes, and developer reactions in one product feed.
Follow this Weaviate release-notes page to spot useful features, risky migrations, noisy announcements, and source links before they hit your backlog.
Where Weaviate updates are checked
Changes.Watch links back to official changelog and release-note sources so summaries stay easy to verify.
Related changelog channels
Use channels to follow groups of tools around a stack, workflow, or topic.
Weaviate updates by type
Rolling windows show how many product updates landed in the last 7, 30, 90, and 365 days, grouped by existing changelog semantics.
2
7
33
149
Latest Weaviate changelog updates
Bound digest RPC memtable snapshots to the requested key range and improved async replication performance by negotiating binary digest encoding and carrying byte ID digests end‑to‑end.
Added new REST search endpoints (bm25, hybrid, near object, aggregate counts) and backup role inclusion feature.
Fix async replication handling, including local resolution of replicas, leaked goroutine fixes, and proper tenant restoration on failures.
Introduced Namespaces with control‑plane isolation, suspend handling, local roles, and API‑key masking for multi‑tenant clusters
Added batch flat ContainsAny/ContainsAll/ContainsNone to the inverted index with optional runtime gate.
Fix read‑repair bug to never use a search result as payload and resolve several panic and shutdown issues across replicas, indexing, and LSMKV
Fixed read‑repair to never use a search result as payload, preventing repair errors
Added a new REST near‑text search endpoint and unified background monitoring metrics;
Eliminate per-call allocations in search SelectProperties for better performance
Accelerated BM25 and LSM read paths with cursor tombstone handling, tiered merged filters, lock‑free tombstone reads and memtable snapshot caching, cutting search latency.
Fixed multiple concurrency and LSM store issues, including batch vectorization deadlocks, WAL sorting, and tombstone handling;
Added crash‑atomic in‑place segment cleanup for LSM KV and sorted WALs before recovery to ensure correct memtable activation
Improved async replication stability and performance: local replica resolution, digest mode to skip full value copies, reduced scheduler workers, and fixes for goroutine leaks, read‑repair, and hash‑tree scan issues.
Stabilized async replication with buffer allocation per hash beat, batched hashtree pre‑filtering, and race‑free shutdown handling.
Hardens async replication and read‑repair pipelines, fixing hashtree initialization races, serialization issues, and ensuring proper handling of concurrent writes.
Improve queue robustness by recovering from corrupt chunk headers, salvaging torn sealed chunks, and adding validation of chunk metadata.
Added automaxprocs support for cgroup v2 (backport to v1.36)
Optimized BM25 scoring loops, data structures, and query setup for significant performance gains
Added new generative‑deepseek module and expanded vector modules (OpenAI, Google, AWS) with endpoint, location, and dimensions settings.
Introduced the generative‑deepseek module and added endpoint, location, and stop settings to OpenAI, Google, and AWS text2vec modules.
Add generative DeepSeek module with new stop and location settings, plus support for dimensions in AWS text2vec and location in Google text2vec modules
Auto‑enable async replication when effective replication factor=1 and actual replication factor>1
Added a configurable rate limiter to the batch simple module
Disable debug endpoints by default and add validation for default vector index runtime settings and replication factor bounds at startup.
Introduced a rate limiter for batch simple logic in modules.
GA release of HFresh with extensive performance, memory and allocation optimizations and improved async error handling
Add rate limiter logic to batch simple processing
Compaction aborts instantly on cancel, fixing lsmkv behavior.
Compaction now aborts immediately when cancelled, improving responsiveness.
Compaction aborts immediately on cancel and async delete operations are now idempotent for missing sources, improving DB reliability.
Increase SSB memory limit threshold to 90% for better stability
Raise SSB memory limit threshold from 80% to 90%
Added release.sh script with driver and enhanced schema validation for reserved property suffixes;
Added new text2vec-digitalocean module for vectorization
Add new text2vec-digitalocean module for text vectorization via DigitalOcean
Added basic auth support to Raft cluster endpoints.
Fixed backup GCS module to use delimiter listing, avoiding full object scans.
Added async replication improvements: CompareDigests mode, corrected frequency handling, and fixed a race condition in the scheduler dispatch logic.
Added async replication scheduler, per‑level discriminants for hash‑tree peer comparison, and basic‑auth support to Raft cluster endpoints.
Fixed various stability issues across replication, RAFT, HNSW, and Object TTL, including preventing shutdown hangs and race conditions.
Disable backup test compactions to reduce flakiness in CI
Speed up collection export snapshots by enabling concurrent tenant de‑activation.
Fix secondary index size accumulation for varying secondary key counts
Persist compression metadata after preloading vectors to avoid data loss on restart.
Fixed secondary index size accumulation for varying key counts and refactored bucket creation to use options for secondary indices.
Fix secondary index size accumulation for varying key counts
Introduced Secure MCP Server (preview) API for AI agents plus internal gRPC communication hardening and async replication optimizations.
Added baseURL validation, descriptive downgrade errors, Google AI Studio API‑Key support, and improved HNSW sparse visited lists with compactv2 downgrade handling.
Added backup and restore support for inactive tenants, including validation of base backup IDs and minor code improvements.
Added on-demand query profiling, sparse HNSW visited‑list support, queue backup, and transferred queue metrics ownership to the scheduler.
Added baseURL validation and enhanced backward‑compatibility error handling for module downgrades and lazy‑load shard disabling
Added token source authentication for GCS backups and improved backup list speed; introduced name logging for CycleManager.
Added Gemini Embedding 2 multimodal and audio support via Google AI Studio models to the multi2vec‑google module.
Added token source authentication for GCS backups and improved backup list speed and reliability.
Added audio support and Gemini Embedding 2 multimodal model to the multi2vec‑google module, plus configurable DisableDimensionMetrics and default sharding count env var.
Added Google AI Studio model support (including Gemini Embedding 2 multimodal) and audio handling to the multi2vec‑google module, plus a DisableDimensionMetrics config option.
Added support for Google AI Studio models in the multi2vec‑google module and accompanying tests.
Updated default propagation settings for async replication.
Fixed BM25 search to retain nil values, correcting query behavior.
Updated default propagation settings for async replication.
Add file‑based incremental backup support and a metric for total streams opened in SSB
Add file‑based incremental backup support
Introduced file‑based incremental backups and delayed tombstone cleanup until cache is prefilled
Delay quantization until after the cache is prefilled, fixing cache-related issues.
Delay quantization until after cache prefill to fix cache-related quantization errors.
Delay quantization until after the cache is prefilled
Fix replication retry behavior and correct error handling for nil shards
Introduced non‑blocking segment deletions feature.
Backport Claude's settings to fix configuration issues
Introduce non‑blocking segment deletions to improve LSM store cleanup
Server‑side batching is now GA, adding smarter acknowledgments, memory‑aware allocation checks, and improved error handling and throughput.
Fixed geo filtering after data re‑ingestion and corrected rerank hybrid result handling.
Fixed tombstone cleanup, heap usage during compaction, and multiple race conditions (e.g., atomic dimensions, OS churn) across the system.
Fixed several bugs: API key self‑deletion, Geo filtering after re‑ingest, thread‑wide os.Chdir removal, lint errors, and test flakiness.
Fixed empty flat index and removed unsafe os.Chdir that affected all threads.
Improve compaction memory usage and reduce heap consumption during set compaction
Added context parameters to LSM object methods and made time‑based deletion the default strategy for new collections
Added context parameter to putObjectLSM and mutableMergeObjectLSM methods (refactor).
Fix race condition in RQ compression and prevent symlink loops during vectors folder migration.
Add async replication improvements and related async features
Fix symlink loop issue when compressing vectors folders
Refactor backup acceptance tests for all backends
Refactored backup acceptance tests to cover all backends.
Add per-node backup size to usage metrics
Fixed proper loading of docId during WAL recovery
Move memtable node creation out of lock and reduce allocations in neighbor connections for better performance
Move memtable node creation out of lock and reduce allocations in neighbor connections to improve concurrency.
Moved memtable node creation and parts handling outside locks, reducing allocations and contention
Reduce network traffic by sending object merge vectors across the cluster only when necessary
Add bucket load limiter to improve load handling
Add a bucket load limiter to improve per‑bucket load handling
Added a bucket load limiter to improve request handling stability.
Added a bucket load limiter and introduced video modality support for the multi2vec‑voyageai module.
Added a bucket load limiter to improve request handling performance.
Precalculates PQ distance look‑up table with branch‑less logic and reduces context check frequency for faster searches
Fix NaN avgPropLength bug causing invalid average property length calculations
Improve PQ distance calculation
Fixed order by lowest ID for tied scores
Fix order-by behavior to use the lowest ID when scores are tied
Fix runtime configuration to be applied immediately on startup, including module configs
Fix runtime config loading to apply overridden settings immediately on start and correct module config loading
Added naive batch processing for text2vec and multi2vec modules and a maintenance‑mode interceptor for the gRPC server.
Fixed runtime configuration to load overridden settings immediately on startup.
Load overridden runtime configuration immediately at startup.
Added Object TTL feature with status endpoint, search filters, and automatic index timestamp handling.
Increase test tick to avoid flakiness and update OIDC testcontainers compose settings
Improved lazy segment lock contention and introduced http.Protocols handling
Added maintenance mode interceptor for gRPC, BaseURL support for Cohere’s reranker, and updated OIDC settings prefix; introduced http.Protocols in server configuration.
Improve lazy segment lock contention and add extra safety checks to prevent panics.
Prefix OIDC runtime settings with "authentication" instead of "exp"
Reduced lock contention during shard loading by using RWLock.
Use RWLock to reduce lock contention when loading shards
Reduce lock contention during shard loading by using an RWLock
Fixed several bugs: swapped sync.WaitGroup for ErrorGroupWrapper in workers, corrected Gemini API batch requests, handled empty batch objects, and improved download worker safety with temporary files.
Fixed Gemini API batch request handling and ensured empty objects are ignored in batch text vectorizer.
Fixed Gemini API batch request implementation
Added debug endpoints for dynamic gomemlimit and max CPU settings, and introduced an internal gRPC server mirroring the REST cluster API
Add debug endpoints to dynamically adjust gomemlimit and max CPUs, and introduce zstd compression for backups while removing backup chunksize
Introduced replication scaling features and updated replication scale URL with collection and factor parameters.
Implemented replication scaling plan with URL parameters and added batch API support for the text2vec-google module and dimensions setting for Cohere modules.
Added replication scaling plan, URL enhancements, and configurable Geo HNSW index settings.
Improved scratch space cleanup to prevent leftover temporary data
Enhance WriteTo method to ensure scratch space is cleaned up and synced after writing
Fix resource leak by deferring shard.release in loops and adding gocritic.deferInLoop rule
Introduced separation of advertise and bind addresses in memberlist raft networking to support multi‑data‑center clusters.
Always defer waitgroup in raft apply to ensure proper cleanup.
Fixed backup path handling and related resource leaks
Introduced separation of advertise and bind addresses for Raft memberlist to enable multi‑data‑center deployments
Reworked LSM store foreground/background synchronization to prevent lock contention and latency spikes.
Fix a data race in the commit logger and other concurrency bugs (waitgroup defers, queue shutdown).
Fixed data races, waitgroup handling, and various bugs in Raft, backup, replication, and usage modules.
Added backup list sorting (asc/desc) with backup size output
Fix data race in commit logger (PR #9443)
- Fixed multiple Raft transport and RPC issues, improving memberlist config and preventing self‑join and rollout errors in the cluster.
Add 1% sampled queries to the slow log and improve compression logging for vectors
Fixed multiple Raft communication and transport issues to avoid lost writes, self‑join loops, and rollout errors, and improved backup shard locking.
Fixed race conditions, panics, and migration bugs in vector indexing and backup shard locking.
Raft transport and node discovery fixes improve stability and avoid invalid ports and self‑join issues.
Added new logging and observability features: wand slow‑log with context cancellation, 1 % sampled queries in slow log, class/shard/vector IDs in HNSW logs, and stack traces in cycle manager.
GA release of Collection Aliases with backup/restore support, case‑insensitive names, and alias‑aware RBAC and API operations
Fixed multiple replication and backup issues, including lost writes, alias recovery errors, and async replication unlocking errors.
Fixed hybrid search grouping, final result handling, and case‑sensitive schema deletion in Raft, reducing crashes and improving search accuracy.
Fix replication metrics in coordinator and read repair steps, and remove unused metric
Added wand slow‑log support, context cancellation, and new reasoningEffort/verbosity settings for OpenAI generative module;
v1.32.8 - LSM store segments read performance improvement, OpenAI GPT-5 reasoning parameters support
Fixed numerous bugs such as validation errors, alias resolution, Raft joiner status checks, schema panic, and commit‑log reliability.
Added reasoningEffort and verbosity params to OpenAI generative module and maxTokens support plus model name/temperature tweaks to AWS generative modules.
Updated OpenAPI docs to list backups.
Added multi2vec‑aws and text2vec‑morph modules, including support for Amazon Nova models.
Added support for Amazon Nova models in modules
Added dimensions setting and switched default model to Gemini embedding 001 in the text2vec-google module (default dims 768).
Fix deletion of inactive tenants and resolve handling of broken compaction segments
Fixed tenant deletion to also remove inactive tenants
Fix async replication deadlock and introduce async queue retries for robustness
Added support for OpenAI GPT‑5 model in the generative Google module
Fixed async replication deadlock by pausing compaction during ObjectDigestsInRange and added async queue retry logic
Adjust vectorizer modules validation logic in named vectors feature.
Fixed infinite loop in LSM bucket cursor by switching to iterative inverted cursors
Added dimensions support for text2vec modules and enriched usage metrics (RQ bits, schema, vector dimension tracking)
Added OIDC group claim parsing from string to support token group mapping
Reduce allocations in segment buffered reader and add filterable/rangeable optimizations for faster queries
Added read‑only RBAC role, custom OIDC JWKS URL support, and multiple usage‑module enhancements (shard jitter, dynamic index fields, object count metadata, cold‑tenant handling).
Added read‑only built‑in RBAC role and support for custom OIDC JWKS URLs
Fixed multiple bugs: async replication thread‑safe shuffler, BMW filtered search infinite loop, hybrid index panic, backup cancellation handling, and viewer groups removal at startup
Fix panic in hybrid index BQ/PQ implementation.
Added class‑creation limits and propagated last‑update timestamps during object replication
Added new read‑only RBAC role, Jina v4 embeddings support, and OIDC audit‑log configuration with AWS IAM certificate download.
Added ability to read segment files containing extra information.
Expose list‑backups endpoint and fix GCS backup listing; improve async replication handling and make hybrid search offsets consistent.
Introduced Collection Alias (preview) with full CRUD endpoint, GQL schema support and RBAC handling
Added Gemini embedding model support to the text2vec‑google module.
Add Gemini embedding model support to the text2vec‑google module and rename the text2colbert module to text2multivec.
Added Gemini embedding model support to the text2vec-google module.
Added domain to RBAC audit logs and introduced lazy segment loading.
Enhanced RBAC audit logs with source IP, domain, and viewer groups; added static-key imports and runtime String() support for dynamic values
Allow imports with static keys and add RBAC enhancements such as viewer groups and source IP logging; introduce runtime support for String() method on dynamic values.
Added support for OIDC self‑signed certificates.
Added support for OIDC self‑signed certificates
Fixed multiple replica‑movement bugs, import cycle issues, async replication handling, and AWS SageMaker integration
Ensure fully loaded segments are always read from memory
Introduced Cost‑Aware Sort with a new query planner and inverted‑index sorter, delivering ~200× faster sorting.
Add several new metrics for Raft (FSM index, single node recovery, last applied index) and include MD5 headers for S3 requests
Fix flaky test caused by unnormalized vectors
Introduce a raft timeout multiplier and adjust Query() and Apply() retry logic (new raft config feature).
Fix raft snapshot structure compatibility and add metrics for index tracking, node recovery, and last applied index.
Added Raft FSM index metrics, single-node recovery metric, and improved WAL reuse and log level handling
Added a metric to track the last applied index on startup
Improved LSM store and BM25 performance, including better average property length handling and reduced writes when creating segments
Updated HuggingFace and Mistral vectorizer modules to support the latest APIs.
- Updated HuggingFace, Mistral, VoyageAI, and Cohere vectorizer modules to support the latest APIs and removed model name validation
Introduced optimized mmap package and related performance improvements, including optimized bloom filter writes and reduced segment write overhead
Reduce segment write overhead and simplify index writes when no secondary indices are present
Fixed Google module end-to-end tests (PR #8062).
Simplifies writing indices when no secondary indices are present
Fixed raft bug where lastAppliedIndex wasn't updated on successful operations.
Enhanced RBAC snapshot support with recovery tests, downgrade path, and nil‑handling fixes.
Fixed missing memwatch handling in segment group segment
Add named vectors to GroupHitAdditional struct
Optimized low‑level I/O: segment‑index header parsing, commit logger, bloom filter reads, and introduced an optimized mmap package.
Optimized header reading to precompute compaction
Introduced optimized mmap package for improved memory mapping performance.
BM25 block enhancements: configurable segment inspection limit via env var, skip searches with empty allowList, and fix filter‑matching bug.
Improved LSM store performance with tokenizer caching, concurrent cache handling, and optimizations for segment index parsing, commit logger writes, bloom filter reading, and Mappair byte creation.
Fixed several issues: erase empty WAL files, async replication digest fetching, BM25 filter matching condition, and made RAFT trailing logs configurable.
Introduced configurable Blockmax migrator options (collections/properties/tenants, shard reload after reindex, rollback selection) and new BM25 block capabilities (higher segment inspection limit, skip empty allowList, trigger reindex vi...
Fix: erase empty WAL files to prevent stale state.
Add an arm64‑only option to the Docker push script and rename Docker tags to .arm64/.amd64.
Fixed concurrency and deadlock bugs in vectorization batching, BM25 segment handling, HNSW tombstone cleanup, and queue chunk recovery.
Introduces a runtime config manager for live updates of settings (e.g., collection limit, autoschema, async replication) without restarts
Added arm64‑only option to Docker push script and updated Docker tag naming conventions
Fixed several replication and concurrency bugs including async indexing deadlock, batch insert concurrency limits, RAFT command handling, and recovery WAL logging.
Fixed BlockMax tombstone handling and global IDF computation
- Fixed multiple stability issues including backup subdirectory handling, shard initialization races, raft schema ops, and tombstone handling for BlockMax vectors.
Added binary quantization using bit shifts, split image fields in dynamic provider, and made model optional in GenerativeOpenAI
Added concurrency tests for RAFT schema operations
Implemented binary quantization using bit‑shifts and added Text2vec‑Weaviate authentication support.
store node id alongside its address to improve raft bootstrapping logging
Added concurrency limit to shard loading and enabled concurrent writes/search in rangeable memtables
Added configurable concurrency limit for shard loading and cycle manager to reduce goroutine usage.
Added concurrency limit to shard loading and prevented excess catching‑up log output
Introduce a concurrency limit for shard loading to improve stability and performance.
Added ability to configure the number of goroutine workers used by the cycle manager.
Made the number of goroutines used by the cycle manager configurable.
Added configurable number of goroutines for the cycle manager via a new setting
RBAC becomes GA with extensive permission enhancements, group management, new env vars, silent internal authorizations, and updated documentation.
Fixed RBAC batch permission handling, adding a custom matcher for collection vs tenant checks and relaxing permission requirements for batching.
Fixed metrics handling for schema restores and corrected replica warning messages.
Improved schema restore metrics and added concurrent DB update handling with GQL rebuild on reload;
Add timeout limits to Docker push step and fix double import of proto API package.
Added new NVIDIA vectorizer modules (reranker, generative, text2vec, multi2vec) and expanded gRPC aggregate capabilities.
Fixed allowlist closure and mergeDocIDs handling for single‑child cases
Added schema metrics (collections and shard counts) and reduced schema cluster traffic by limiting get class calls to the leader
Added schema‑related metrics (shards count, schema collections, telemetry collections) and a flag to disable async replication
Added schema‑related metrics (collections count, schema collections, and per‑node shard count) and telemetry for collections
Increase default gRPC max message size to 100 MB
Raise default gRPC max message size to 100 MB and update the gRPC health probe version.
Add cluster communication metrics and tighten RBAC enforcement (auto‑tenant gating, permission validation, env var unification).
Fixed hcLogLogrus logger merging, tenant activity unlocking, and backup test context handling.
Add metrics for cluster communication
Added Ollama batch endpoint support and updated CI pipelines, including removal of the Anthropic e2e test and switch from Trivy to Orca scanning
Upgrade golang.org/x/crypto to fix CVE‑2024‑45337 (security)
Upgrade golang.org/x/crypto to fix CVE‑2024‑45337 (security).
Introduce VoyageAI multimodal module, add support for X‑Goog‑* headers and Ollama batch endpoint
Added VoyageAI multimodal module and Ollama batch endpoint support
Add VoyageAI multimodal module
Fixed redundant retries for non‑multi‑tenant queries and corrected infinite preload progress logging
- Fixed LSMKV compaction/flush synchronization, redundant retry logic, infinite preload logs, and flaky async indexing tests.
Added Weaviate vector DB header to JinaAI client.
Introduced the new multi2vec‑jinaai module for Jina AI vector generation.
Add multi2vec‑jinaai vectorizer module
Add multi2vec‑jinaai vectorizer module
- Reduced test flakiness and fixed shard initialization issues (adjusted sleep times, bucket load handling).
Fixed NaN handling and propLength defaults, corrected bytebuf behavior, and adjusted segment cleanup test sleep times.
Added a new multi2vec‑cohere module and expanded multi‑vector capabilities, including OpenAI default model update and JinaAI client extraction.
Fixed NaN handling and propLength defaults, and corrected byte buffer handling to avoid nil slices.
Added new multi2vec‑cohere module and expanded default modules list with text2vec‑mistral and reranker‑jinaai
Added dynamic backup locations and a new multi2vec‑cohere module, plus range‑filter migration fixes.
Implemented parallel prefill for compressed vector cache, non‑blocking segment drop in compaction, and made GRPC message size configurable (2 GB default) with an HNSW visited‑list pool limit.
Added parallel prefill for compressed vector cache and non‑blocking segment drop in compaction, boosting HNSW performance and reliability.
Return raw error body when decoding fails and fix a data race in vectorizer tests
Added option to limit the HNSW visited‑list pool and cancel timed‑out HNSW searches
Add validation to prevent class creation with invalid replication factor and enforce batch object class names, ref2vec config, and near‑object parameters.
Fixed backup-related issues: updated Swagger description, improved error handling and cancellation responses, and added validation for replication factors in multi‑tenant mode.
Added LSM segment cleanup feature with progress indicator for schema catch‑up.
Added ACORN‑based minority filters for HNSW and introduced multi‑target vector search allowing multiple inputs per target vector
Introduced LSM store segments cleanup feature.
Fixed long locking during tombstone cleanup, BM25 vector return, and ensured single‑node Raft recovery doesn't modify the DB.
Fixed BM25 search vector return bug and ensured single‑node Raft recovery does not modify the DB
Fixes long locking issue when cleaning up tombstones.
Added support for Jina AI V3 embeddings and new VoyageAI models.
Added backup list and cancel endpoints with status validation and release handling for cancellations.
Added backup list and cancel endpoints