@lemmaoracle/parser

Natural language query parser for the Lemma protocol. Converts natural language into structured attribute queries using on-device LLM inference via Transformers.js.

Privacy-preserving: All parsing happens client-side. No query text is sent to any server.

Install

pnpm add @lemmaoracle/parser

Usage

import { initParser, parseNaturalQuery, cleanup } from "@lemmaoracle/parser";

// 1. Initialize (downloads model on first run)
await initParser();

// 2. Parse natural language → structured query
const query = await parseNaturalQuery("users over 18 in Japan");
// → { attributes: [{ name: "age", operator: "gt", value: 18 }, { name: "country", operator: "eq", value: "Japan" }] }

// 3. Use with @lemmaoracle/sdk
import { create, attributes } from "@lemmaoracle/sdk";
const client = create({ apiBase: "https://api.lemmaoracle.com" });
const results = await attributes.query(client, query);

// 4. Cleanup when done
await cleanup();

API

initParser(modelId?, progressCallback?)

Initialize the parser with an optional custom model and progress callback.

• modelId — ONNX model ID (default: onnx-community/Qwen3-0.6B-ONNX)
• progressCallback — Callback for model download progress

parseNaturalQuery(query)

Parse a natural language string into a ParsedQuery object.

cleanup()

Release model resources.

Types

type ParsedQuery = {
  attributes: AttributeCondition[];
  targets?: { schemas?: string[] };
  proof?: { required: boolean; type?: "zk-snark" | "opaque" };
};

type AttributeCondition = {
  name: string;
  operator: "eq" | "neq" | "gt" | "gte" | "lt" | "lte" | "in" | "contains";
  value: string | number | boolean | (string | number)[];
};

Whitepaper

See §4.10 — Verified Attributes Query.