A vector database in a single file.

import recern_vector as rv

db = rv.Database.open_or_create("docs.rvec")
docs = db.create_collection("docs", dim=384, metric="cosine")
docs.upsert_many(ids, embeddings, metadatas)

hits = docs.search(query, k=5, filter={"lang": "en"})
docs.explain(query, k=5)  # strategy, visited nodes, time

$ recern-vector query docs.rvec docs --like doc-42 -k 3 \
    --filter '{"lang": "en", "year": {"$gte": 2020}}' \
    --explain

  #  id          distance  metadata
  1  doc-42      -0.00000  {"lang":"en","year":2024}
  2  doc-18861    0.63155  {"lang":"en","year":2022}
  3  doc-10899    0.63834  {"lang":"en","year":2024}

  strategy      hnsw
  ef            64
  selectivity   ~19.9% of records match the filter
  visited       7,244 nodes
  distances     7,330 computed
  time          529 µs

Python bindings take NumPy float32 arrays without converting each element. Batch inserts build the index on all cores.

• Portable Python, Rust and Node APIs

  Use the same collection, record and filtered-search operations on embedded Recern, Recern HTTP and Qdrant. Profile v1 preserves logical IDs, metadata precision, distance semantics and explicit backend capabilities.

• Verified migration

  Export a consistent snapshot from embedded Recern or Recern HTTP and import into new Recern or Qdrant collections. Includes checksums, dry-run, resume and read-back verification. Int8 exports contain stored, dequantized vectors.

• HNSW and exact search

  A graph index with per-collection m, ef_construction and ef_search, and an exact scan when you need ground truth.

• Cosine, L2 and dot product

  Each collection has a fixed dimension and metric. Cosine vectors are normalized when they are stored.

• Metadata filters

  Equality, $in, ranges and nested $and, $or and $not. A cost planner samples filter selectivity, chooses exact or HNSW search and adapts the search effort.

• Atomic embedded inserts

  Embedded Recern builds batch indexes on all cores and rejects an invalid batch before writing. Qdrant batches do not carry an atomicity guarantee.

• Upsert, delete and compact

  Replace or delete records by id, then compact to rebuild the collection without deleted records.

• Introspection API

  stats(), explain() and estimate_recall() are part of the API, not a separate tool.

• Python bindings

  A PyO3 wheel for CPython 3.11 and later, with type hints.

• Command-line tool

  init, create-collection, insert, query, inspect, recall, checkpoint and compact.

• Stable format 2

  Versioned files, checksums and golden compatibility tests, with existing format-1 snapshots still readable.

• WAL and checkpoint

  save() commits changed pages to the WAL; checkpoint() writes a self-contained snapshot. Interrupted trailing records are ignored during recovery.

• Optional int8

  Per-collection int8 codes with per-vector scales. Measure recall on your data to choose the storage and accuracy tradeoff.

• Native and portable Node.js

  Node-API 8, TypeScript and portable async operations on native workers. Verified npm tarballs for Linux x64/ARM64, macOS Intel/Apple silicon and Windows x64 are in the GitHub release.

Recern Vector 0.4.0 · 2026-11-02 · embedded f32 and int8; HTTP and Qdrant are outside this measurement.

1. Available

   0.4.0

   • Single-node HTTP server with authentication and durable writes
   • Python, Rust and Node: embedded Recern, Recern HTTP and Qdrant
   • Migration and snapshots tested; metadata precision preserved
   • Verified packages and native builds for five platforms
   Release packages and compatibility
2. Planned

   0.5.0

   • Incremental persistence and automatic checkpoints
   • Public npm packages with platform selection
   • Reproducible f32/int8 memory and recall benchmarks
3. Planned

   1.0.0

   • Stable API and deprecation policy
   • Upgrade, backup and migration suites
   • Operational limits and compatibility matrix

Embedded data must fit in memory; saves still encode the full image. Migration requires an immutable export and exclusive access to new destination collections. Consistent Qdrant export is not supported. Clustering and a hosted service are outside the 1.0 scope.