HeliosDB Nano v3.11.0 Release Notes

Release Date: 2026-04-16 Theme: RAG-native primitives — graph traversal, hybrid search, and MCP extensions

Highlights

This release introduces a complete set of RAG (Retrieval-Augmented Generation) primitives directly into Nano’s embedded engine. You can now do graph traversal, full-text BM25 search, vector similarity search, and rank-fusion — all from SQL, in a single binary, with no external services.

What’s New

Graph Engine (NEW)

A lightweight in-memory graph layer with adjacency lists, exposed through SQL functions:

-- Traverse from a starting node, depth 2 across 'KNOWS' edges
SELECT * FROM graph_traverse(
  start_node => 'uuid-here',
  edge_label => 'KNOWS',
  direction  => 'OUT',
  depth      => 2
);

-- Shortest path with Dijkstra
SELECT * FROM graph_shortest_path(
  from_node => 'uuid-a',
  to_node   => 'uuid-b',
  algorithm => 'dijkstra'
);

Algorithms supported: BFS, Dijkstra, bidirectional BFS.

Full-Text BM25 + Hybrid Search (NEW)

-- Pure BM25 keyword search
SELECT * FROM documents
WHERE bm25_match(content, 'graph database for AI')
ORDER BY bm25_score(content, 'graph database for AI') DESC
LIMIT 10;

-- Hybrid: BM25 + vector with RRF fusion
SELECT * FROM hybrid_search(
  table_name   => 'documents',
  query_text   => 'graph database for AI',
  query_vector => embed('graph database for AI'),
  vector_column=> 'embedding',
  text_column  => 'content',
  fusion       => 'rrf',
  limit        => 20
);

Fusion strategies: RRF (Reciprocal Rank Fusion), MMR (Maximal Marginal Relevance), Linear.

Compiled Query Plans (NEW)

Frequently executed queries can be cached as compiled plans, skipping parser/planner overhead:

PREPARE COMPILED find_similar AS
  SELECT * FROM hybrid_search(
    table_name   => 'documents',
    query_vector => $1,
    vector_column=> 'embedding',
    text_column  => 'content',
    fusion       => 'rrf',
    limit        => 20
  );

EXECUTE find_similar(embed('my query'));

Plans are cached in an LRU and survive across connections within a process.

MCP Server Extensions (NEW)

6 new tools and 2 new resources for AI agents using Model Context Protocol:

Tools:

heliosdb_bm25_index — create BM25 indexes
heliosdb_hybrid_search — BM25+HNSW with RRF/MMR
heliosdb_graph_add_edge — add graph edges
heliosdb_graph_traverse — traverse from a node
heliosdb_graph_path — find shortest path
heliosdb_embed_and_store — vectorize text and INSERT in one call

Resources:

heliosdb://schema/{table} — schema introspection
heliosdb://stats/{table} — table statistics

Performance: Arena Allocation (Internal)

A new RequestArena (built on bumpalo::Bump) is now available throughout the engine for per-request allocations. Hot paths in graph traversal and search use it to reduce allocator pressure under load. Application code generally does not need to interact with it directly.

Known Limitations

Graph and BM25 storage are in-memory. Persistence to RocksDB column families is planned for v3.12.0. For embedded use cases that rebuild the index on startup, this is fine; for long-lived deployments, persist source data to tables and warm graph/BM25 indexes on startup.
MCP extensions ship in mcp_extensions module. The legacy src/mcp/ server module is being migrated; in the meantime, new tools are accessed through the extensions module.

Migration

No breaking changes. v3.11.0 is fully wire-compatible with v3.10.0.

To use the new SQL functions, no schema migration is required. The first bm25_match or graph_traverse call materializes the index on demand.

Compatibility Matrix

Component	Version
PostgreSQL wire	14, 15, 16
MySQL wire	5.7, 8.0
Oracle wire	(subset)
MCP protocol	1.0