LoginSign Up
← Back to all sparks
M

Manticore Search

DEVOPS
Velocity7.5

Open-source database for full-text search and analytics

manticoresearch.com ↗

Manticore 29.9 ships chunked multi-vector embeddings and mmap column access, closing gaps with dedicated vector DBs.

searchvector-searchembeddingsopen-sourcerag
◆Current state
Manticoresearch is releasing at high cadence, shipping major capabilities alongside a stream of correctness fixes. The 29.9.0 release consolidates chunked auto-embeddings with multiple strategies (mean, fixed, recursive, sentence), float_vector_array for multi-vector document storage, mmap-based columnar attribute access, and AWS credential-chain backup authentication — all in a single open-source artifact. The 29.8.x series concurrently fixed hybrid search correctness, Elasticsearch-compatible bulk error handling, and RT table embedding metadata.
◆Where it's heading
The engine is systematically replacing external dependencies for AI workloads. Native chunking means no upstream text-splitting service, auto-embeddings with configurable input limits means no external embedding pipeline, and float_vector_array means no separate vector database for chunk-level retrieval. Manticore is positioning as the single system that ingests, chunks, embeds, and searches — a self-hosted alternative to a Qdrant or Weaviate stack that requires orchestrating multiple services. The cloud-aware backup additions suggest it's also targeting managed deployments.
◆Prediction
The hybrid search correctness fixes in 29.8.x reveal active work on BM25+KNN fusion. The next likely move is a configurable retrieval reranker or a scoring blend API that lets applications tune the balance between lexical and vector relevance without writing fusion code themselves.

◆Recent moves

  1. 1d ago

    29.9.3: Buddy dependency bump

    A Buddy dependency version bump with no user-visible behavior change — routine maintenance in the 29.9 release series.

    View source ↗
  2. 4d ago

    Manticore Search 29.9.0

    ⚡ SPARK

    Manticore 29.9 delivers chunked auto-embeddings, float_vector_array for multi-vector document storage, mmap columnar attributes, and AWS credential-chain backups — a feature cluster that together removes the need for separate chunking, embedding, and vector storage services in a self-hosted stack.

    View source ↗
  3. 6d ago

    29.8.4: fix: apply hybrid weight filters after fusion

    Applying hybrid weight filters after fusion rather than before fixes a correctness issue that silently returned wrong results for queries combining full-text weight expressions with vector search — this is the kind of correctness bug that erodes trust in hybrid retrieval.

    View source ↗
  4. 6d ago

    29.8.3: fix: restore Buddy fallback for bulk item errors

    Restoring Buddy's AutoSchema fallback in bulk error handling — an internal compatibility fix ensuring NDJSON processing works correctly across the HTTP layer.

    View source ↗
  5. 8d ago

    29.8.1: fix: align /_bulk item error responses

    Fixing Elasticsearch-compatible /_bulk to return HTTP 200 for processed batches while reporting per-item errors correctly prevents clients like Fluent Bit from retrying already-applied writes — a correctness fix with real operational impact for teams using standard log shippers.

    View source ↗
  6. 9d ago

    29.8.0: S3 backup gains AWS credential provider chain

    Backup 1.11.0 enables the AWS SDK credential provider chain for S3 backups — IRSA, shared credentials, ECS task roles, and EC2 instance profiles now work without static key configuration, which is how production AWS deployments manage credentials.

    View source ↗