@forgrit/design-intelligence

Embedding-based retrieval, hybrid scoring, RAG pipeline, and intelligence primitives for @forgrit/design-engine. The ML/LLM layer that turns design intent into ranked recipe selection.

Status: early-access (v0.x). Pre-1.0 releases may include breaking changes in minor bumps until the API stabilizes at v1.0.0.

Node 20+. Hard deps: zod + @forgrit/design-engine@^0.1.2. Optional peer dep: onnxruntime-node (only if you want real MiniLM embeddings; package works without it via TF-IDF fallback).

What's in the box

Layered intelligence primitives for design-recipe selection:

TF-IDF + Validation

• TfidfEngine, tokenize — text-similarity retrieval
• ResponsiveIntentValidator — constraint-penalty checks for responsive design intent

Embedding + Retrieval

• MiniLMEmbedder — MiniLM-L6 embeddings via optional onnxruntime-node peer
• RecipeEmbeddingCache — persistent embedding cache for design recipes
• DualRetriever — hybrid retrieval combining TF-IDF + embeddings
• reciprocalRankFusion — RRF score combination for ranked lists

Scoring

• HybridRecipeScorer + HybridRecipeSelector — weighted multi-signal recipe ranking
• HybridScoringConfigSchema — Zod-validated config + getDefaultConfig / getFallbackConfig

Repair + LLM

• SchemaRepairer — heal partial/broken Zod-shaped JSON from LLMs
• LocalLLMRuntime — local node-llama-cpp runtime adapter (optional)
• Planner — design-intent planning via LLM with confidence scoring
• ConfidencePolicy — disagreement stats + fallback decisions

Versioning + Models

• ManifestStore, buildManifest, diffManifests — versioned intelligence pipelines
• ModelManager, MODEL_REGISTRY — model lifecycle + health

Memory + Critic

• IntentMemory — design intent profile persistence
• DesignCritic — dimensional critique + suggested-swap output

Training + Eval

• FineTuningDatasetBuilder + QLoRATrainer — training pipeline primitives
• Stage2EvalRunner, Stage3GraduationRunner, Stage4GraduationRunner — evaluation graduation gates
• STAGE2_EXTENDED_PROMPTS, STAGE3_EXTENDED_PROMPTS — extended eval datasets

RAG + Review + Monitoring + Improvement + A/B

• RAGPipeline — full RAG with re-ranking
• HumanReviewSampler — human-review batch sampling
• ProductionMonitor — alerts + latency tracking
• ContinuousImprovementPipeline — closed-loop signal-to-config improvement
• ABFramework — A/B trial orchestration + reporting

What it's NOT

• Not a standalone design engine. Depends on @forgrit/design-engine for recipe data, semantics dictionary, and rule-based intelligence.
• Not a model server. The LLM runtime is in-process via optional node-llama-cpp.
• Not a hosted vector DB. Embeddings cache to local files; bring your own scale-out store.

Install

npm install @forgrit/design-intelligence @forgrit/design-engine
# or
pnpm add @forgrit/design-intelligence @forgrit/design-engine

For real embedding inference (optional):

npm install onnxruntime-node

For local LLM runtime (optional):

npm install node-llama-cpp

Quick start

import { TfidfEngine, tokenize } from '@forgrit/design-intelligence';
import { DESIGN_RECIPES } from '@forgrit/design-engine';

// 1. Build a TF-IDF index over the design recipes
const engine = new TfidfEngine();
for (const recipe of DESIGN_RECIPES) {
  engine.addDocument(recipe.id, recipe.description ?? recipe.id);
}

// 2. Query
const results = engine.search('admin dashboard with charts', { topK: 5 });
console.log(results); // [{ id, score }, ...]

Versioning

0.1.x is early-access. The public API may evolve before 1.0.0 locks semver.

License

MIT — see LICENSE.

Links

• npm: https://www.npmjs.com/package/@forgrit/design-intelligence
• Source: https://github.com/forgrit-ai/forgrit/tree/main/packages/design-intelligence
• Issues: https://github.com/forgrit-ai/forgrit/issues
• ForGrit: https://forgrit.ai

Sibling packages

• @forgrit/blueprint — app-spec format
• @forgrit/llm-cost — credit ledger primitives
• @forgrit/design-engine — recipe + theme combinatorial engine