Mini PDF Q&A (Next.js + Pinecone + Gemini)

An end‑to‑end example that lets you:

• Upload a PDF, extract its text, chunk it, embed with Gemini, and store in Pinecone
• Ask a question, retrieve the most relevant chunks from Pinecone, and get a grounded answer with Gemini

Tech stack

• Next.js (App Router)
• Pinecone JS SDK
• Google Generative AI SDK (Gemini) for embeddings and answers
• pdf-parse for PDF text extraction

Prerequisites

• Node.js 18+
• Pinecone account with an index created
• Google AI Studio API key (Gemini)

Environment variables Create a .env.local file in the project root:

GEMINI_API_KEY=your_gemini_api_key
PINECONE_API_KEY=your_pinecone_api_key
PINECONE_INDEX=your_pinecone_index_name

Important: Your Pinecone index dimension must match your embedding model. This project uses Gemini model text-embedding-004 which is 768 dimensions. Ensure your index is created with dimension 768.

Install and run

npm install
npm run dev

Open http://localhost:3000

How it works

• Upload flow (POST /api/upload)
  • Receives a PDF via multipart form-data
  • Extracts text with pdf-parse
  • Splits into ~500‑character chunks
  • Generates embeddings via Gemini text-embedding-004
  • Upserts vectors to Pinecone with text stored in metadata
• Ask flow (POST /api/ask)
  • Embeds the question
  • Queries Pinecone (topK=3, includeMetadata=true)
  • Concatenates matched chunks as context
  • Calls Gemini (gemini-2.5-flash) to answer using the context

Endpoints

• POST /api/upload
  • Body: multipart form-data with field "file" (application/pdf)
  • Response: { message: string }
• POST /api/ask
  • Body: { "question": string }
  • Response: { "answer": string, "context": PineconeMatch[] }

Quick API examples

# Upload a PDF
curl -F "file=@./sample.pdf" http://localhost:3000/api/upload

# Ask a question
curl -X POST http://localhost:3000/api/ask \
  -H "Content-Type: application/json" \
  -d '{"question":"What does the document say about X?"}'

Configuration knobs

• Chunk size: src/app/api/upload/route.ts (regex for chunking)
• Retrieval topK: src/app/api/ask/route.ts
• Models: src/lib/embeddings.ts (text-embedding-004) and src/app/api/ask/route.ts (gemini-2.5-flash)

Project structure

src/
  app/
    api/
      upload/route.ts   # Upload + index PDF chunks to Pinecone
      ask/route.ts      # Retrieve + answer with Gemini
    page.tsx            # Simple UI for upload + ask
    globals.css         # Theme + Tailwind config
  lib/
    embeddings.ts       # Gemini embeddings (text-embedding-004)
    pdf.ts              # PDF text extraction
  services/
    pinecone.ts         # Pinecone client

Troubleshooting

• PineconeBadRequestError: Vector dimension 0 or mismatch
  • Ensure chunks are non-empty
  • Ensure embeddings.ts returns a non-empty vector array
  • Confirm your Pinecone index is created with dimension 768
• 401/permission errors
  • Verify GEMINI_API_KEY and PINECONE_API_KEY are set and valid

Deployment

• Vercel works great: set env vars in project settings
• Or run locally in production mode:

npm run build
npm start