Changelog

Features 🎶

• #10750 - BM25 scoring directly on the full-text payload index [docs]
• Milestone [#57] - TurboQuant: new turbo8 and turbo16 vector datatypes [docs]
• #11125 - Per-element payload indexes: new nested index parameter, faster nested condition filtering. [docs]
• #11180, #11263 - New code tokenizer for full-text index [docs]
• #10466 - Configurable audit log output: audit.output writes audit events to file (default), stdout or stderr [docs]
• #10992 - Collection creation time: created_at is recorded and returned in collection info [docs]
• #10711 - Substring filter condition: match: { substring: "..." } (full-scan only) [docs]
• #10231, #10287, #10296 - New formula query expressions: acosh, max, min [docs] and [docs]
• #10597, #10598 - Configurable ID tracker memory placement [docs]

Improvements 🎵

• #11009, #11020 - Faster HNSW payload links build - builder sized to the block, not the segment, and reused
• #11083 - Faster HNSW healing - payload blocks reuse old graph links; m = 0 graphs can heal too
• #11166 - Smaller HNSW upper layers - floor() in level assignment matches the HNSW paper (6.25% of points on level 1 at m = 16, not 25%)
• #10976 - Faster PQ training - k-means sample gathered in one pass, not one per chunk
• #10975 - Faster scalar quantization search - AVX512-VNNI and AVX-VNNI kernels for int8 dot product
• #10977 - Less search slowdown during optimizations - the whole optimization job runs at low thread priority
• #11078, #11080 - Fewer cancelled optimizations - optimizers are recreated only when their config changes
• #11234 - Faster point retrieval - ids resolved in one batch per segment
• #11070 - Faster snapshot transfers and downloads - 4 MiB stream buffers instead of 4 KiB
• #11253 - Faster shard transfers - transfer_raw_points and transfer_raw_payloads on by default
• #11206 - Safer snapshots - unique snapshot names, and a stored snapshot is never overwritten
• #11221 - No stuck snapshot downloads - 60s timeout when the server never responds
• #11163 - No blocking thread leak on telemetry timeout - the timeout also applies inside the blocking task
• #11164 - Clearer tar unpack errors - release builds report the error kind (e.g. timed out) instead of "malformed archive"
• #11118 - Log timestamps respect timezone - set with the TZ environment variable, UTC by default
• #10235 - timeout parameter for the get points API
• Milestone [#58] - Less disk usage with hnsw_config.inline_storage - vectors are stored only inside the HNSW graph file, without a separate vector storage copy
• #10790 - Faster enable_hnsw change on payload indexes - updated in place instead of dropping and rebuilding the index
• #10906 - Faster search over on-disk vectors - batches of vectors are prefetched with madvise(WILLNEED) before scoring
• #10907 - ACORN on m = 0 graphs - search uses the payload-block links of the graph
• #11255 - Better shard distribution in small clusters - auto-sharded collections get 2 shards per node in 3+ node clusters while collections are fewer than nodes
• #11273 - No internal request stalls behind NAT and load balancers - HTTP/2 keepalive on internal p2p gRPC
• #11289 - Configurable snapshot storage timeouts and retries - new timeout_sec, connect_timeout_sec, max_retries, retry_timeout_sec in S3, GCS and Azure config

Bug fixes 🎼

• #10932, #10971 - Fix collection reads stalling while a shard transfer holds the shard holder lock
• #10949 - Fix live-lock between shard transfer start and consensus operations
• #10956 - Fix collection being blocked while a shard snapshot is extracted
• #11075 - Fix must_not conditions being ignored when equal to an indexed must condition
• #11101 - Fix datetime filters with microseconds behaving differently with and without a payload index
• #10948 - Fix renaming an alias to an existing collection name being accepted
• #10887 - Fix transfer staying registered after crash replay of replica removal during transfer abort
• #11121 - Fix S3 snapshot upload completing truncated on a local read error
• #10796 - Fix Japanese tokenizer ignoring min_token_len, max_token_len, stemmer, ascii_folding and case-insensitive stopwords
• #10754 - Fix quantization build with MSVC-like compilers (clang-cl)

Security 🔑

• #10854 - Internal auth is enforced by default: service.enforce_internal_auth now defaults to true when an API key is configured
• #10718 - Startup warning when distributed mode has an API key configured but enforce_internal_auth is off
• #11275, #11290 - Reject vector names that escape the segment directory, including . and ..
• #11278 - Update rustls-webpki and webpki-root-certs for RUSTSEC-2026-0382 and RUSTSEC-2026-0383

Qdrant Edge 📀

• #11085 - WAL-less mode for edge shards: EdgeConfig::wal_enabled(false) skips WAL writes, updates are durable after flush()
• #11084 - Fix updates lost on crash: unacknowledged WAL operations are replayed when an edge shard is loaded
• #11082 - Fix EdgeShard::flush not acknowledging the persisted version in the WAL