Manticore Search 29.9.0: Chunked auto-embeddings and mmap columnar access

Manticore Search 29.9.0 has been released. The largest change is in auto-embeddings: long documents can now be split into searchable chunks, and a document can keep several vectors instead of compressing all of its content into one. This release also makes mmap the default access mode for columnar attributes, adds safer controls for embedding workloads, UTF-8 identifiers, better backup support, and fixes across hybrid search, KNN, bulk ingestion, grouped search, and schema changes.

This post covers everything shipped after 29.0.2, from 29.0.3 through 29.9.0.

❤️ We’d like to thank @tudorvasinca for their work on PR #4857 , PR #4859 , and PR #4873 .

Upgrade notes #

There is no release-wide mandatory data migration. Existing tables and configuration can be upgraded normally, but a few fixes need follow-up if the earlier behavior already affected your data:

• German sharp-s normalization is opt-in. Switching an existing table to lemmatize_de_v2 orlemmatize_de_v2_all changes indexed terms, so rebuild a plain table or replay documents into a new RT table.
• Authenticated BACKUP introduces thebackup authorization action. Before downgrading to 29.3.12 or earlier, remove anybackup grants.
• Columnar attributes now use mmap by default instead of bufferedfile reads. Setaccess_columnar_attrs='file' explicitly if you need to retain the previous access mode.

Long documents can keep more than one embedding #

The old auto-embedding path produced one vector per document and truncated text that did not fit the model input window. That works for titles and short descriptions, but it means a relevant paragraph near the end of a long article may never reach the index.

Manticore Search now supports five chunking strategies :

• truncate keeps the previous behavior and remains the default.
• mean embeds every chunk and averages the results into one vector.
• fixed splits text into fixed-size token windows.
• recursive prefers paragraph, line, sentence, and space boundaries in that order.
• sentence groups complete sentences up to the configured limit.

The multi-vector strategies use float_vector_array . Each chunk competes independently during KNN search, but Manticore returns the document once and uses its closest chunk for knn_dist():

CREATE TABLE articles (
  title text,
  content text,
  chunks float_vector_array knn_type='hnsw' hnsw_similarity='cosine'
    model_name='Xenova/all-MiniLM-L6-v2' from='title,content'
    chunk_strategy='sentence' max_tokens='256' overlap_tokens='32'
);

INSERT INTO articles (id, title, content)
VALUES (1, 'Rotating certificates', 'A long guide with many sections ...');

SELECT id, knn_dist()
FROM articles
WHERE knn(chunks, 5, 'how do I rotate a certificate');

MAX_TOKENS, OVERLAP_TOKENS, and MAX_CHUNKS control chunk size, shared context at boundaries, and the maximum vector count. Declare a model-backed float_vector_array when creating the table: adding one later with ALTER TABLE ... ADD COLUMN and rebuilding its embeddings are not supported yet.

Put a ceiling on local embedding work #

Long-context models can make a single large input unexpectedly expensive, especially on CPU. The new MAX_INPUT_TOKENS column option caps how much of each input is sent to a local embedding model:

ALTER TABLE articles
MODIFY COLUMN chunks MAX_INPUT_TOKENS='512';

The change applies to embeddings generated afterward; existing vectors stay as they are. Set it during CREATE TABLE or change it later without re-embedding the table. A value of 0, or leaving the option out, uses the model's own limit.

This release also fixes two less visible problems around this path. Models no longer remain cached after an embedding column is modified, and configurations with different API_TIMEOUT or MAX_INPUT_TOKENS values no longer collide and reuse the wrong cached model.

UTF-8 identifiers and better German matching #

Table, field, and attribute names now follow one consistent safe UTF-8 identifier syntax . RT, percolate, distributed, template, and plain tables can use localized names such as Chinese or Cyrillic identifiers across DDL, expressions, field selectors, and inferred source schemas.

German AOT morphology also gains opt-in sharp-s normalization. With charset_table=non_cont,german and morphology=lemmatize_de_v2 or lemmatize_de_v2_all, forms such as Straße, Strasse, and STRAẞE match in ordinary whole-word searches. If index_exact_words=1 is enabled, exact-word queries can still distinguish the ß and ss forms.

Better load testing and backups #

manticore-load can now benchmark through the HTTP JSON API with --http. Writes use /bulk, searches use /search, and --table selects the target table. Its reports now include local searchd RSS during the run plus peak RSS, disk, and CPU statistics at the end, including aggregate monitoring for multi-command workloads.

Manticore Backup now works with authenticated Manticore Search installations through username/password or bearer-token credentials. SQL BACKUP has a dedicated authorization action and checks read access to the selected tables.

S3 backups and restores can also use the AWS SDK credential provider chain when static keys are not set. That includes IRSA, shared credentials, ECS task roles, and EC2 instance profiles. Temporary credentials can supply AWS_SESSION_TOKEN.

Columnar attributes use mmap by default #

Manticore Search now defaults access_columnar_attrs to mmap. The operating system maps and caches *.spc columnar-attribute files on demand, without prereading the whole file at startup. The previous buffered path remains available with access_columnar_attrs='file'.

ALTER TABLE also reopens replaced columnar storage with the configured access mode, so an altered table no longer falls back to a different reader than the one requested.

Vector and grouped-search fixes #

Several fixes target queries that were valid but could return incomplete results or fail under a particular table layout:

• Distributed and sharded KNN queries with a local shard no longer rescore merged 1-bit-quantized results twice, which could crash the coordinator or return the wrong nearest neighbor. (Issue #4791 )
• KNN queries with additional filters avoid a redundant knn_dist prefilter when HNSW already excludes documents without vectors. (PR #4861 )
• LENGTH() on afloat_vector_array now reports the number of vectors rather than its internal storage-word count. (PR #4879 )
• Hybrid search with GROUP BY retains all buckets, including MVA groups, and follows both final and within-group ordering. (Issue #4639 )
• Hybrid filters on weight() and expressions or aliases derived from it now run after fusion against the final text weight instead of being ignored. Weight-dependent filters insideOR trees remain unsupported and return an explicit error. (Issue #4889 )
• Multi-chunk RT grouping no longer risks duplicate groups, incorrect split counts, or a hang while ordering COUNT(DISTINCT ...) results. (Issue #4856 )
• Document-ID filters whose signed representation is negative now follow the lookup index's unsigned ordering. (Issue #4774 )

There are crash fixes here too: a second hybrid-search statement in a multi-statement request (PR #4864 ), distributed JSON aggregation sorted by a string attribute (Issue #4822 ), and dropping a table during auto-embedding precommit (Issue #4860 ) are all handled safely now.

Bulk ingestion behaves predictably #

Elasticsearch-compatible /_bulk requests now return HTTP 200 once a batch has been processed, while individual failures remain visible through errors: true and per-item statuses. Duplicate create actions return per-item 409 version_conflict_engine_exception errors, including duplicates within the same batch. This prevents clients such as Fluent Bit from retrying writes that already succeeded.

Fixed-length gzip-compressed /bulk bodies are also decoded correctly when they arrive across multiple socket reads. And if native bulk processing fails because the target table does not exist, the request can again reach Manticore's auto-schema fallback with valid NDJSON while preserving the expected bulk response envelope.

More reliability fixes #

The rest of the release closes a broad set of operational and compatibility problems:

• searchd --stopwait no longer hangs while a sharded table is being rebalanced after a node rejoins. (Issue #3905 )
• Compatible older binlogs replay safely during upgrade and completed RT chunks are published before clean shutdown. (Issue #4808 ) Fatal replay diagnostics also name the relevant recovery flag. (Issue #4811 )
• indexer creates missing parent directories for plain-table paths when the nearest existing parent is writable. (Issue #4793 )
• UUID document IDs no longer cause stored text fields to come back empty. (Issue #4833 )
• ALTER TABLE ... RENAME preserves hidden remote embedding API keys without exposing them inSHOW CREATE TABLE . (Issue #4842 )
• Sequel Ace 5.3.1+ compatibility probes work again. (Issue #4828 )
• JSON /search keeps distances for negatedNEAR and proximity operators. (Issue #4784 )
• Internal string-sort helper columns no longer leak from LEFT JOIN results. (Issue #4788 )
• Malformed binary API SEARCH element counts are rejected instead of terminatingsearchd . (PR #4790 )

For the complete list, see the Version 29.9.0 changelog .

Get Manticore Search 29.9.0 #

Install or upgrade Manticore Search with the installation guide . Review the upgrade notes above if you use columnar attributes, German AOT morphology, or authenticated backups.

Need help or want to connect? #

• Join our Slack
• Visit the Forum
• Report issues or suggest features on GitHub
• Email us at [email protected]