Vertrez DocsBot

Small local RAG demo. Upload PDFs / markdown / text files, ask questions, get answers grounded in the documents you uploaded, with the source passages cited under each answer.

Embeddings run locally (Transformers.js, MiniLM-L6). Answer generation goes to Gemini's free tier. Everything stays on your machine — there's no deployment, no database.

Setup

npm install
cp .env.example .env.local   # paste your Gemini key
npm run dev

Free Gemini key from https://aistudio.google.com — no credit card needed.

Open http://localhost:3000, drop a file in the sidebar, ask away.

Optional: pre-seed docs

If you want the bot to start with some baseline knowledge that survives restarts, drop .md or .txt files into data/seed-docs/ and run:

npm run seed

That builds data/seed-embeddings.json, which the server loads on startup. Re-run any time you change a seed file.

Without seeding, the bot just starts empty — uploads added through the UI live in memory until the server restarts.

Stack

• Next.js 15 (App Router) + TypeScript
• Tailwind v3
• @huggingface/transformers — local embeddings (Xenova/all-MiniLM-L6-v2)
• @google/genai — gemini-2.5-flash for generation
• In-memory vector store with cosine search
• pdf-parse for PDFs

Scripts

npm run dev      # dev server
npm run build    # production build
npm run start    # run the production build
npm run seed     # rebuild seed-embeddings.json from data/seed-docs/

Notes

• Uploads are in-memory only and gone on server restart. For persistence, swap lib/vectorStore.ts for a real vector DB.
• Generation lives behind a single generateAnswer() in lib/generate.ts — switching providers is a one-file change.
• The model cache (~25 MB) is downloaded once on first run and cached locally.
• Guardrails in lib/config.ts: 1000 char question limit, 5 MB upload limit, 50 chunks per upload, top-5 retrieval.