VelesDB v1.12.0

[1.12.0] - 2026-04-05

Added

• Cross-collection MATCH queries (Issue #495) — @collection annotation on MATCH node patterns
  enables cross-collection graph queries. Syntax: MATCH (p:Product@products)-[:STORED_IN]->(inv:Inventory@inventory).
  Results are enriched with payloads from annotated collections (alias-prefixed fields).
• MATCH via /query endpoint — MATCH queries can now be executed via Database::execute_query
  using _collection parameter or SELECT ... FROM <collection> WHERE MATCH ... syntax.
  Previously, MATCH was rejected at the Database level.
• Cross-type JOIN tests — VectorCollection JOIN MetadataCollection validated with BDD tests.
• Graph API parity — 7 new REST endpoints for complete graph operations:
  • DELETE /collections/{name}/graph/edges/{id} — remove edge by ID
  • GET /collections/{name}/graph/edges/count — total edge count
  • GET /collections/{name}/graph/nodes — list all node IDs
  • GET /collections/{name}/graph/nodes/{id}/edges — node edges with direction filter (in/out/both)
  • GET/PUT /collections/{name}/graph/nodes/{id}/payload — get/upsert node payload
  • POST /collections/{name}/graph/traverse/parallel — multi-source BFS traversal
  • POST /collections/{name}/graph/search — embedding similarity search on graph nodes
• CLI graph commands — graph remove-edge, graph count, graph search
• REPL graph commands — .graph remove-edge, .graph count, .graph search, .graph store-payload, .graph get-payload, .graph nodes with full help documentation
• Core — GraphCollection::traverse_bfs_parallel() for multi-source BFS with deduplication
• OpenAPI — all new graph endpoints registered in API documentation
• Tests — 238 BDD tests, 4447 lib tests, 13 new server tests

Performance

• Bitmap pre-filter for filtered search (#487) — adaptive strategy selection based on
  real selectivity: full-scan brute-force for ≤1% selectivity, HNSW+bitmap for 1-30%,
  post-filter fallback for >30%. Eliminates massive over-fetching on selective filters
• CSR graph traversal v2 (#491) — CsrSnapshot with edge IDs + labels, ArcSwap lock-free
  adjacency, EdgePredicate pushdown (290ns for label-filtered BFS vs 3.4µs unfiltered = 12x),
  lazy CSR rebuild on read instead of every mutation
• Bulk insert v2 pipeline (#488) — DirectVectorWriter bypasses ShardedVectors overhead,
  AsyncIndexBuilder with background thread for deferred HNSW construction
• ComponentScores optimization (#476) — SmallVec<[(&'static str, f32); 4]> eliminates per-result heap allocation
• Bitmap NEQ support — Neq conditions now use universe bitmap subtraction
• Secondary index backfill — create_index scans existing payloads to populate the index
• LIKE→BM25 fallback — metadata-only LIKE queries use BM25 text index for candidate narrowing
• Native batch edge loading — add_edges_batch with single lock acquisition cycle (1M+ edges/s)
• 19 functions CC > 8 reduced to ≤ 8 — all non-SIMD functions now comply with Codacy CC ≤ 8

Fixed

• BFS dedup — CSR and EdgeStore BFS no longer produce duplicate results for nodes
  reachable via multiple paths (diamond graph fix).
• DISTINCT in early-return paths (#475) — SELECT DISTINCT now applied in NOT-similarity and union query paths.
• NEAR + MATCH + metadata filter (#474) — co-occurring metadata filters no longer silently dropped in hybrid search.
• list_indexes includes secondary indexes — was only returning property/range indexes.
• rrf_k propagated to hybrid_search_with_filter (#472) — was hardcoded to 60.
• CSR label filter — edges with unresolvable labels now excluded when rel_type filter is active.
• Cross-collection enrichment — logs tracing::warn when @collection references a non-existent collection.
• AsyncIndexBuilder drain in flush — flush() and flush_full() now drain the AIB buffer into HNSW.
• Tauri drop_index — uses require_vector_collection instead of deprecated require_collection.
• LangChain filter= kwarg — backward-compatible alias restored alongside metadata_filter=.

Changed (Breaking)

• BFS cycle behavior — BFS no longer re-emits already-visited nodes when a cycle closes.
  Code relying on duplicate entries for cycle detection must be updated.
• ComponentScores type — changed from SmallVec<[(String, f32); 4]> to SmallVec<[(&'static str, f32); 4]>.
  External code constructing SearchResult with custom component scores must use &'static str literals.
• Python relationship_types= alias (#490) — traverse_bfs/dfs now accept both rel_types= and relationship_types=.
• CLI commands restructured — flat commands (velesdb list) grouped into sub-commands (velesdb collection list).
• License — added Attribution clause for public-facing applications (visible link
  to velesdb.com required, Enterprise license waives requirement).
• Rate limit — increased default from 100 to 100K QPS for local-first deployment.

What's Changed

• perf(hnsw): recall at 100K scale + panic→Result + test fixes by @cyberlife-coder in #494
• license: add Attribution clause for public-facing applications by @cyberlife-coder in #496
• feat: add image search tutorial (Hamming + CLIP two-pass pipeline) by @cyberlife-coder in #506
• refactor: align tutorial naming with article (Bouncer/Detective) by @cyberlife-coder in #507
• Release: develop → main (perf optimizations + graph API parity + CI updates) by @cyberlife-coder in #497

Full Changelog: v1.11.0...v1.12.0