Framework generators — React, Vue, Angular, Svelte, HTML
Component registry — 502 curated snippets (357 component + 85 animation + 60 backend)
ML quality scoring — Embeddings, quality validation, anti-generic rules
Feedback system — Self-learning, pattern promotion, feedback-boosted search
Template compositions — Pre-built page templates with quality gating
Brand integration — Transform branding-mcp tokens into design context
LLM providers — Ollama, OpenAI, Anthropic, Gemini with auto-fallback

Installation

npm install @forgespace/siza-gen

Lightweight Entry (`/lite`)

A zero-native-dependency entry point for edge runtimes (Cloudflare Workers, Deno, Bun). Provides context assembly without the registry/database/ML stack.

import { assembleContext } from '@forgespace/siza-gen/lite';

const ctx = assembleContext({
  framework: 'react',
  componentLibrary: 'shadcn',
  tokenBudget: 4000,
});
// ctx.systemPrompt — ready to use as LLM system prompt

43 KB vs 1.87 MB full bundle. Includes brandToDesignContext, designContextStore, and all core types.

Usage

import {
  searchComponents,
  initializeRegistry,
  GeneratorFactory,
} from '@forgespace/siza-gen';

await initializeRegistry();
const results = searchComponents('hero section');
const generator = GeneratorFactory.create('react');

What's inside

Module	Description
`generators/`	React, Vue, Angular, Svelte, HTML code generators
`registry/`	502 snippets — 357 component + 85 animation + 60 backend
`ml/`	Embeddings (all-MiniLM-L6-v2), quality scoring, training pipeline
`feedback/`	Self-learning loop, pattern promotion, feedback-boosted search
`quality/`	Anti-generic rules, diversity tracking
`artifacts/`	Generated artifact storage and learning loop

LLM Providers

Built-in multi-provider support with auto-fallback:

import { createProviderWithFallback } from '@forgespace/siza-gen';

// Tries Ollama first (local), falls back to OpenAI/Anthropic/Gemini
const provider = await createProviderWithFallback();

Supports: Ollama (local), OpenAI, Anthropic, Gemini (via OpenAI adapter).

Brand Integration

Transform branding-mcp tokens into design context:

import { brandToDesignContext } from '@forgespace/siza-gen';

const designContext = brandToDesignContext(brandIdentity);

Python ML Sidecar

An optional Python FastAPI sidecar handles compute-intensive ML operations. When unavailable, the system gracefully degrades to Transformers.js and heuristics.

cd python && pip install -e ".[dev]"
python -m uvicorn siza_ml.app:app --port 8100

Or via npm:

npm run sidecar:start     # Launch Python sidecar
npm run sidecar:test      # Run Python tests (41 tests)

Endpoint	Description
`POST /embed`	Sentence-transformer embeddings
`POST /embed/batch`	Batch embeddings
`POST /vector/search`	FAISS k-NN similarity search
`POST /score`	LLM-based quality scoring
`POST /enhance`	LLM-based prompt enhancement
`POST /train/start`	LoRA fine-tuning via PEFT
`GET /health`	Liveness check
`GET /metrics/report`	ML observability metrics

Fallback chain: Python sidecar → Transformers.js/local LLM → heuristics.

Development

npm install && npm run build
npm test                  # 424 tests, 21 suites
npm run validate          # lint + format + typecheck + test
npm run registry:stats    # Report snippet counts