Break~ish change Latest Images - Preview of 1.0.12

[orneryd]

Hey all!

I needed to de-pollute the properties namespace for nodes so I moved the managed embedding information to a node struct property instead. This means you can use any property field name in Properties for the node and it won't get clobbered by anything. If you use the managed embeddings, you will probably need to re-generate them entirely or risk tomb-stoning your existing embeddings in the data files.

I also moved the e2e search response time down to 7-8 ms for the entire e2e RAG retreival pipeline including embedding the string query (most production systems target <~100ms for the entire pipeline). This was tested on a dataset of 300K nodes of content translated between multiple languages (resulted in ~940K embeddings.)

IVF-HNSW scaling for 1024 dims overhead for the cluster routing breaks even somewhere between 40M-50M embeddings. For all size datasets, we auto-select strategy based on the size (tunable still) but defaults have changes based on the new search implementation.

Added

• Per-database configuration and routing improvements

  • Added per-database config override support and supporting server/dbconfig plumbing.
  • Expanded multi-database handling across API, search, and UI flows.

• RBAC and access-control primitives

  • Added/expanded role, entitlement, privilege, allowlist, and database-access components in pkg/auth.
  • Extended RBAC enforcement for gRPC and related server surfaces.

• Search/index persistence capabilities

  • Added append-only vector file persistence and index versioning helpers.
  • Added BM25 v2 implementation plus persistence and benchmark coverage.
  • Added search build settings and performance/profile test coverage.

• Heimdall integration enhancements

  • Added OpenAI/Ollama remote generator support and OpenAI-compatible action naming.
  • Added MCP context/tool exposure and improved plugin integration patterns.

Changed

• Search architecture and performance

  • Refactored search startup/build lifecycle for responsiveness.
  • Hardened persistence/restore for HNSW/IVF-HNSW, with improved readiness/reporting.
  • Improved BM25 indexing throughput via batching and runtime defaults.
  • Added non-blocking index persistence paths and better build observability.

• Config model consolidation

  • Unified config handling across server/embedding/cypher/search layers.
  • Reduced config-side mutations of user data and cleaned up defaults.

• Cypher and query execution behavior

  • Improved transactional semantics and added regression coverage.
  • Improved SET/map/reduce evaluation behavior and error messaging.
  • Expanded parser compatibility for more complex expressions/workflows.

• macOS File Indexer and app UX

  • Refined tagging semantics (label-safe behavior), indexing reliability, and UI affordances.
  • Improved keychain prompt/window focus behavior and installer/default presets.

Fixed

• Restored vector text-query behavior under per-database executor/embedder wiring.
• Fixed Cypher parsing and relationship/traversal edge cases in complex queries.
• Fixed parameter handling in toLower, optional-match, and distinct-related paths.
• Fixed snapshot retention/atomic WAL test behavior and related durability checks.
• Fixed startup/bootstrap auth/write-path regressions in server + macOS flow.
• Fixed large embedding transaction guardrails and model-display issues.

Docs & Developer Experience

• Major docs refresh across architecture, security/RBAC, features, performance, and user guides.
• Added/updated guidance for clustering, hybrid/vector search, Heimdall agentic loop, and MCP tooling.
• Build/ops updates including Windows target fixes and cleanup of deprecated loader paths.

---

Thanks for your bug reports and suggestions! please keep them coming ❤️