Loka is in Developer Preview. Core engine, SPARQL+, vector indexing, and HTTP server are fully functional. Here's what's done and what's next.
The loka install-agent command sets up a database interactively for AI agents. Currently debugging path handling, HNSW index rebuild on startup, and error messaging for non-interactive environments. Goal: any AI agent says "set up a database" and it just works.
Cardinality estimation and predicate pushdown are done. Remaining: planner should choose HNSW index scan vs SPO triple scan based on cost, and adaptive execution should observe intermediate result sizes at runtime and reorder mid-query.
HNSW topology is exposed as virtual RDF triples (loka:hnswNeighbor). Remaining: make property path evaluation produce correct ANN results by letting greedy descent + beam search emerge from the graph structure.
Standard SPARQL property paths can't express "traverse and stop when a condition is met." The UNTIL extension adds per-step predicate evaluation during traversal with backtracking support and ordered traversal semantics.
During low-usage periods, rebuild HNSW indexes and rediscover pseudo-tables in the background. Old indexes stay live until the new ones are ready, then atomic swap.
The semantic web ecosystem runs on Java. The SDK code is complete; publishing to Maven Central requires GPG signing configuration, Sonatype account setup, and pom.xml updates for distribution management. This is the highest-priority SDK to publish.
Python (PyPI), TypeScript (npm), Rust (crates.io), Go (pkg.go.dev), and .NET (NuGet). All SDK code is complete and tested. Publishing requires registry account setup for each platform. See SDK Accounts Setup.
SPO/POS/OSP indexes over interned u64 IDs. sled LSM-tree persistence with crash recovery. RDF-star quoted triple support. Bulk insert at 20K triples/sec.
Per-predicate HNSW indexes with cosine, Euclidean, and dot product distance metrics. SIMD acceleration (AVX2/SSE). Tombstone-based deletion. Multiple entry points for search quality.
Full SPARQL 1.1 parser, cost-based query planner, and iterator-based executor. Extensions: VECTOR_SIMILAR, VECTOR_SCORE, metric-specific search operators. Property paths, subqueries, aggregates, all working.
SPARQL endpoint, Graph Store Protocol, REST API for triples and vectors. Content negotiation (JSON, XML, CSV, TSV, Turtle). Optional passcode auth, rate limiting, query timeouts, periodic backups.
Automatic discovery of columnar indexes from graph structure. SIMD-accelerated column scans, zonemap pruning, segment-level storage. Deep subgraph pattern mining with fan-in detection.
Python, TypeScript, Go, Rust, Java, .NET. All with client-side OWL validation, typed error handling, and vector search support. Code complete, awaiting registry publishing.
Desktop and web client. Graph visualization, SPARQL editor, HNSW health dashboard, triple editor, OWL ontology viewer, backup management. 90% feature complete.
6 GitHub Actions workflows. Clippy + rustfmt enforcement. 256 tests, 50+ Criterion benchmarks tracked historically. Docker image. Cross-platform install scripts.
loka update command, --version flag, startup version check, and HNSW rebuild endpoint. The CLI can check for new releases, self-update in place, and trigger index rebuilds after version upgrades.
Dual-mode MCP server supporting both serverless and server configurations. Provides 8 maintenance tools for AI agents to manage Loka instances through the Model Context Protocol.
Atomic transactions, startup verification, durability guarantees, and isolation. All write operations are fully transactional with crash recovery support.
Cypher and GQL (ISO 39075) as translation layers over SPARQL. The database still speaks SPARQL+ internally; these wrappers let users query with familiar graph query syntax. SQL and MQL are deliberately excluded.
RBAC, encryption at rest, TLS, audit logging, replication, clustering, multi-tenancy. These will be shaped by customer feedback. The open-source core is designed to be sufficient for most use cases.