frontend-agent

Demo - GitHub Pages

A generic, English web/front-end agent fine-tuned from LiquidAI LFM2.5 (230M and 350M), small enough to run entirely in the browser (edge / local inference via wllama, no server-side model). It doesn't just chat - it calls real tools to act on a page (browse, look up, navigate, manage a cart, check out) and answers questions grounded in retrieved context.

Trained for patterns, not lexical knowledge. The weights hold behaviors - tool selection, argument and id binding, positional and quantity resolution, RAG-grounded answering, graceful refusal, and steering back when asked something off-topic - not any specific catalog or domain facts. Meaning and grounding are supplied at runtime: each turn the host injects a compact context - the on-screen items (with ids and prices), the cart, and any retrieved knowledge - and the model grounds strictly in what it is given. Swap the store, swap the injected context - the same model works, no retraining.

Client library - frontend-agent is a fully-typed, framework-agnostic package: createFrontendAgent({ persona, view, cart, tools }) returns a Session with a single submit(text) feed point and an event stream. It handles tool-calling with GBNF-constrained decoding (well-formed calls and id-grounding by construction), a configurable Hugging Face model source, and optional RAG adapters.

Demo - demo/ is the reference consumer: three storefronts sharing one on-device assistant. See demo/README.md.

How it works

• Frozen tool roster - a fixed set of eight tools (list_items, get_item, search_knowledge, add_to_cart, remove_from_cart, clear_cart, checkout, navigate) the model learns by name, so a tiny model gets a small, reliable action space. The only per-store variance is the filter set on list_items, which the model reads from the injected schema rather than memorizing.
• RAG as a tool - retrieval is exposed as list_items (catalog) and search_knowledge (guides, policies); the model grounds answers in the results and declines when they don't contain the answer. Backend-agnostic (BM25, vector, hybrid) behind a small adapter - only the result shape is the contract.
• Grammar-guaranteed calls - decoding is constrained to a GBNF grammar, so every tool call is well-formed and every id it emits is one actually present in the injected context - no hallucinated ids.
• Held-out-domain eval - the model is scored on verticals never seen in training; the demo's BrewCraft domain is itself held out, proving the generalization thesis end to end.

Quickstart

npm install
npm run build -w frontend-agent
npm run dev -w demo        # http://localhost:3000

The assistant downloads its GGUF once from Hugging Face (cached in the browser via OPFS) on first activation - no server-side inference.

Licensing

• Code & docs - Apache 2.0
• Model weights - LFM Open License v1.0

Authoring

This repository was substantially authored with the following models, under human direction and review:

• Qwen3 Coder 30B A3B Q4_K_M (local Ollama)
• Claude Opus 4.8 (Anthropic)