Muninn

Muninn is a memory-backed RAG assistant for portfolio sites. It answers from a bundled corpus, streams responses from Groq, and returns citation metadata so every answer can point back to its source.

Named after Odin's raven of memory, Muninn is designed to be small, cheap to run, and easy to deploy.

Features

• Build-time MiniLM embeddings with @xenova/transformers
• Bundled JSON vector index, no hosted vector database
• Groq llama-3.3-70b-versatile streaming responses
• Citation chips via the x-muninn-citations response header
• Optional Upstash Redis rate limiting
• Next.js App Router API route at /api/muninn

Tech Stack

• Next.js 16
• React 19
• Vercel AI SDK
• Groq
• MiniLM local embeddings
• Upstash Ratelimit, optional

Quick Start

npm install
cp .env.example .env.local
npm run build:muninn
npm run dev

Open http://localhost:3000.

Environment

Required:

GROQ_API_KEY=

Optional:

UPSTASH_REDIS_REST_URL=
UPSTASH_REDIS_REST_TOKEN=

If Upstash variables are missing, Muninn still runs locally and skips rate limiting.

Corpus

Muninn's memory lives in data/corpus.json.

After editing the corpus, rebuild the index:

npm run build:muninn

This writes data/muninn-index.json, which is imported directly by the runtime route. The app does not need a database or a vector-store service.

API

POST /api/muninn

Request:

{
  "messages": [
    { "role": "user", "content": "Explain Muninn's RAG architecture." }
  ]
}

Response:

• Body: plain text stream
• Header: x-muninn-citations, URL-encoded JSON citation metadata

Scripts

npm run dev           # start local dev server
npm run build:muninn  # rebuild vector index
npm run build         # build index, then build Next app
npm run start         # start production server
npm run lint          # run ESLint

Deployment

1. Push this folder as its own repository.
2. Add GROQ_API_KEY to the deployment environment.
3. Optionally add Upstash Redis REST credentials.
4. Deploy with the default Next.js build command: npm run build.

License

MIT