LoginSign Up
← Back to all sparks
M

Manticore Search

DEVOPS
Velocity7.5

Open-source database for full-text search and analytics

manticoresearch.com ↗

Manticore Search 29.9 ships multi-vector KNN and chunked embeddings — serious RAG infrastructure now in open source.

searchvector-searchembeddingsopen-sourceragai-native
◆Current state
Manticore Search is releasing rapidly, mixing major capability drops with a tight patch cycle. The 29.9.0 release expanded the AI-search surface significantly: chunked and multi-vector auto-embeddings, float_vector_array for storing multiple KNN vectors per document, configurable embedding input limits for long documents, memory-mapped columnar attributes by default, and authenticated cloud-aware backups. Post-release patches have been stabilizing the new embedding and hybrid search machinery.
◆Where it's heading
Manticore is building toward a full AI-native search layer. The multi-vector storage foundation, chunking strategies, and hybrid search fixes point toward positioning as a credible open-source option in RAG pipeline infrastructure. Operational maturity work — mmap defaults, cloud backups, authentication — is running in parallel with the AI feature surface.
◆Prediction
The next likely additions are query-time re-ranking improvements or richer embedding model integration, building on the float_vector_array and chunking foundation. The hybrid search bug fix streak suggests that area will continue to get attention.

◆Recent moves

  1. 11h ago

    29.11.2: Bump buddy version from 4.4.4 to 4.4.5 (#4926)

    Dependency bump for the buddy component — housekeeping following the 29.11.0 plain table freeze feature, no user-visible effect.

    View source ↗
  2. 14h ago

    Manticore Search adds freeze for plain tables

    Freeze support for plain tables adds an operational control to lock table state — useful for backup consistency and snapshot workflows. Extends the table management surface that 29.9.0's backup authentication work began.

    View source ↗
  3. 3d ago

    29.9.6: fix: prevent deadlock in concurrent BERT auto-embeddings

    Deadlock fix in concurrent BERT auto-embedding requests — a correctness issue in the embedding infrastructure from 29.9.0 that could stall search under parallel load.

    View source ↗
  4. 4d ago

    29.9.5: fix: implicit sorter tdigest leak

    Memory leak in the implicit group sorter during parallel percentile aggregations — a low-visibility fix that matters for long-running analytics workloads with percentile queries.

    View source ↗
  5. 7d ago

    29.9.3: Bump buddy version from 4.4.3 to 4.4.4 (#4901)

    Routine buddy component version bump — dependency maintenance with no user-visible effect.

    View source ↗
  6. 10d ago

    Manticore Search 29.9: multi-vector KNN, chunked embeddings, cloud-aware backups

    ⚡ SPARK

    Manticore 29.9.0 is the largest release in this window: multi-vector storage via float_vector_array, chunked auto-embeddings with configurable strategies, configurable per-column input limits, memory-mapped columnar attributes by default, and cloud-aware authenticated backups. This is the inflection point where Manticore gains the infrastructure to compete in AI-native search.

    View source ↗