@agentix-e/nl2spel-webllm

Browser-local LLM Provider for @agentix-e/nl2spel

GBNF grammar-constrained decoding · 4 optimized models · Zero API cost · WebLLM handles GPU detection natively

API Documentation

→ nl2spel-webllm API

Model Configurations (4)

| Model | Context | VRAM | Quantization | |-------|---------|------|--------------| | Qwen 2.5 1.5B | 32K | 1.5 GB | q4f16_1 | | Gemma 2 2B | 8K | 2.0 GB | q4f16_1 | | Phi-3 Mini 4K | 4K | 2.8 GB | q4f16_1 | | Llama 3.2 3B | 8K | 3.5 GB | q4f16_1 |

Quick Start

import { NL2SpelEngine } from '@agentix-e/nl2spel';
import { WebLLMProvider, recommendModel } from '@agentix-e/nl2spel-webllm';

const engine = new NL2SpelEngine();

// Auto-select the best model for the current device
const modelName = recommendModel(4);  // 4GB VRAM
// → 'qwen2.5-1.5b' (fastest)

const provider = new WebLLMProvider({
  model: modelName,
  onProgress: (p) => console.log(`Loading: ${(p.progress * 100).toFixed(0)}%`),
});

await provider.initialize();
engine.registerProvider(provider);

// Zero API cost, local browser inference
const result = await engine.generate('Amount greater than 1000');

GBNF Grammar Constraints

The WebLLM Provider constrains LLM output via dynamically generated GBNF grammar, ensuring 100% valid SpEL expressions:

import { GBNFGenerator } from '@agentix-e/nl2spel-webllm';

const gen = new GBNFGenerator({ injectContext: true });
const grammar = gen.generate({
  root: { name: 'order', type: 'Order', fields: { amount: { type: 'number' } }, methods: {} },
  variables: { user: { type: 'object' } },
  beans: {},
  types: {},
  functions: {},
});

// grammar output can be directly injected into WebLLM's grammar parameter

Browser Requirements

• WebGPU support (Chrome 113+, Edge 113+, Opera 99+)
• Sufficient VRAM (≥ 1.5 GB)
• HTTPS or localhost environment

License

MIT © 2025 Agentix-E