Chunk Lab
aashi2912/chunk-lab
Visualize como chunking muda resultados de RAG na mesma pergunta
Ver no GitHubPlataforma interativa que demonstra como diferentes estratégias de chunking impactam a qualidade da retrieval em RAG. Executa um pipeline completo: segmentação do documento (fixed-size, recursivo ou sentence-level), embedding com Voyage AI, busca por similaridade e geração de resposta com Claude. Perfeito para entender por que a estratégia de chunking é tão crítica quanto a escolha do banco de vetores.
Instalação
npx convez-ai aashi2912/chunk-lab1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Ensinar estratégias de chunking para RAG
- Comparar impacto de diferentes segmentações em retrieval
- Prototipação de pipelines de embedding e busca
Veredito

Excelente ferramenta educacional que desmitifica RAG através da experimentação prática. O código é limpo, extensível e roda localmente sem atrito — ideal para quem quer entender como chunking afeta retrieval antes de escalar.
README
A tradução para o português deste README ainda está na fila.
Chunk Lab
Same document. Same question. Different chunking — different answer.
An interactive demo showing why chunking strategy matters as much as the vector database you pick. Built to demonstrate end-to-end RAG: real embeddings, real similarity search, real LLM answer generation.
What it does
You give Chunk Lab a document and a question. It runs a complete RAG pipeline:
- Chunks the document using your selected strategy (fixed-size, recursive, or sentence-level)
- Embeds each chunk into a vector using Voyage AI
- Searches by cosine similarity against the embedded query
- Generates an answer with Claude using only the top retrieved chunks
Then you switch strategies and rerun — and see how chunking choice changes everything downstream.
Stack
| Layer | Tech |
|---|---|
| Framework | Next.js 14 (App Router) + TypeScript |
| Embeddings | Voyage AI (voyage-3-lite, 512-dim) |
| Vector store | In-memory cosine similarity |
| LLM | Anthropic Claude (claude-haiku-4-5) |
| Styling | Tailwind CSS |
| Deploy target | Vercel |
The vector store is intentionally in-memory. Pinecone, Weaviate, ChromaDB, and Qdrant all do the same fundamental thing — storage + similarity search. To swap one in, replace lib/vectorStore.ts. The rest of the app stays untouched.
Local setup
1. Clone and install
git clone https://github.com/<your-username>/chunk-lab.git
cd chunk-lab
npm install
2. Get API keys
- Anthropic — console.anthropic.com/settings/keys. Free tier is plenty.
- Voyage AI — dash.voyageai.com/api-keys. Free tier gives 200M tokens/month.
3. Set env vars
Create .env.local in the project root:
ANTHROPIC_API_KEY=sk-ant-...
VOYAGE_API_KEY=pa-...
4. Run
npm run dev
Open http://localhost:3000.
Deploy to Vercel
1. Push to GitHub
git init
git add .
git commit -m "initial commit"
git branch -M main
git remote add origin https://github.com/<your-username>/chunk-lab.git
git push -u origin main
2. Import to Vercel
- Go to vercel.com/new
- Import your
chunk-labrepo - Vercel auto-detects Next.js — no config needed
3. Add env vars in Vercel
In the import flow (or under Project → Settings → Environment Variables), add:
| Name | Value |
|---|---|
ANTHROPIC_API_KEY | your Anthropic key |
VOYAGE_API_KEY | your Voyage key |
4. Deploy
Click Deploy. First build takes ~90 seconds. Subsequent deploys ship in ~30 seconds on every git push.
Project structure
chunk-lab/
├── app/
│ ├── api/chunk/route.ts # The full RAG pipeline endpoint
│ ├── globals.css # Tailwind + base styles
│ ├── layout.tsx # Root layout + metadata
│ └── page.tsx # The UI
├── lib/
│ ├── chunking.ts # 3 chunking strategies
│ ├── embeddings.ts # Voyage AI embedding calls
│ ├── samples.ts # Demo documents
│ └── vectorStore.ts # In-memory cosine similarity
├── .env.example # Template for env vars
├── next.config.js
├── package.json
├── postcss.config.js
├── tailwind.config.js
├── tsconfig.json
└── README.md
Cost notes
For typical usage on the demo:
- Voyage embeddings: ~free. Each run embeds ~10-20 chunks + 1 query. 200M tokens/month free tier is millions of runs.
- Claude: ~$0.001 per run (Haiku is cheap, prompt is small).
If you embed Chunk Lab in a portfolio and it gets 1,000 visits with average 3 runs each → ~$3 in Claude costs. Budget accordingly.
Why no Pinecone or ChromaDB?
The honest answer: for demonstrating chunking effects, an in-memory cosine similarity is identical to what a vector DB does — minus the indexing, persistence, and clustering layers that matter at scale.
Adding Pinecone would have meant: another API key, another account, slower setup, and zero new insight for the demo's purpose. The decision is documented in lib/vectorStore.ts.
If you want to see Chunk Lab v2 with real Pinecone, swap the search() function. The interface stays the same.