Persistent, edge-native knowledge graph storage for agents.

• Portable facade — Unified Client API that works across in-memory, LibSQL/Turso, and Deno KV backends.
• Search — Hybrid retrieval combining keyword FTS5 and vector embeddings.
• Query — Built-in SPARQL engine for declarative graph traversal and reasoning.
• In-memory RDF/JS — Zero-setup N3-based graph store and search for dev, tests, and demos.

deno add jsr:@worlds/client

import { Client } from "@worlds/client";
import { ComunicaSparqlEngine } from "@worlds/client/comunica";
import { RdfjsQuadStore, RdfjsSearchIndex } from "@worlds/client/rdfjs";
import { QueryEngine } from "@comunica/query-sparql-rdfjs-lite";
import { Store } from "n3";

const store = new Store();
const client = new Client({
  quadStore: new RdfjsQuadStore(store),
  searchIndex: new RdfjsSearchIndex(store),
  sparqlEngine: new ComunicaSparqlEngine({
    queryEngine: new QueryEngine(),
    store,
  }),
});

await client.import({
  source: {
    kind: "serialized",
    contentType: "text/turtle",
    data: `@prefix ex: <http://example.org/> .
      ex:Alice ex:bio "Alice explores the depths." ;
               ex:location "Underdark" .`,
  },
});

const searchResults = await client.search({ query: "explores" });
const subject = searchResults.results[0].subject;

const sparqlResponse = await client.sparql({
  query: `SELECT ?property ?object WHERE { <${subject}> ?property ?object }`,
});
console.log(sparqlResponse);

Quad store: Manages RDF triples (subject, predicate, object, graph) with transactional import and export.

Search index: Hybrid retrieval over graph literals, combining keyword FTS5 with vector similarity via an embedding service and quad chunker.

SPARQL engine: Evaluates declarative queries and updates against the graph for structured traversal and reasoning.

Client is the portable facade. Shared modules sit under src/client/:

Regenerate merged API doc JSON with deno task doc:json (writes gitignored docs/api.json). For architecture documentation (package topology, runtime model), see ARCHITECTURE.md. For agent coding rules and conventions, see AGENTS.md.

This package provides the core in-memory RDF/JS backend. Durable backends live in separate packages:

Choosing LibSQL vs Deno KV: LibSQL is the default for hybrid FTS/vector search and faster cold quad index preload at scale. Deno KV can be faster on selective SPARQL execute after preload in long-lived or cached processes. See discussion #69 for benchmark methodology.

import { Client } from "@worlds/client";
import { ComunicaSparqlEngine } from "@worlds/client/comunica";
import { RdfjsQuadStore, RdfjsSearchIndex } from "@worlds/client/rdfjs";
import { QueryEngine } from "@comunica/query-sparql-rdfjs-lite";
import { Store } from "n3";

const store = new Store();
const client = new Client({
  quadStore: new RdfjsQuadStore(store),
  searchIndex: new RdfjsSearchIndex(store),
  sparqlEngine: new ComunicaSparqlEngine({
    queryEngine: new QueryEngine(),
    store,
  }),
});

For LibSQL or Deno KV examples, see the @worlds/libsql and @worlds/denokv repositories.

The agent eval harness lives in a separate repository and runs deterministic assertion checks against a seeded LibSQL world.

Agent integration: search-then-SPARQL two-hop pattern for LLM tool use with hybrid retrieval. See AGENTS.md and ARCHITECTURE.md.

LibSQL/Deno KV benchmarks: Quad index performance methodology, regression policy, and comparison tables live in the adapter repos (@worlds/libsql, @worlds/denokv).

All CI checks must pass before merging updates.

• Documentation
• Wazoo Technologies
• Support

Developed with @wazootech