Paper Mind AI

An intelligent PDF reader & Q&A assistant. Upload a PDF, it identifies the document's domain, builds a lightweight searchable knowledge base, and answers questions using only the content of that document.

Tech stack

Layer	Technology
Backend	Python, Flask
PDF text extraction	pypdf
Retrieval / NLP	scikit-learn — TF-IDF vectorization + cosine similarity to find the passages most relevant to a query (full document is used directly for short PDFs)
Answer generation & domain classification	Anthropic Claude, Google Gemini, or Hugging Face-hosted open models (auto-detected from whichever API key is present)
Frontend	Plain HTML / CSS / JavaScript (no framework) served by Flask, calling the backend's JSON API

There is no client-side call to any AI API — the browser only talks to your own Flask server, which is the one holding the API key.

Setup

cd paper-mind-ai
python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install -r requirements.txt

cp .env.example .env
# then open .env and paste ONE of: Google AI Studio key, Anthropic key, or a
# Hugging Face token (the file must be named exactly ".env")

python app.py

Choosing a provider

The backend checks .env at startup:

• If GOOGLE_API_KEY is set, it uses Gemini (gemini-2.0-flash by default).
• Else if ANTHROPIC_API_KEY is set, it uses Claude (claude-sonnet-5 by default).
• Else if HF_API_KEY is set, it uses Hugging Face (meta-llama/Llama-3.3-70B-Instruct by default).
• If more than one is set, that priority order applies unless you force one with LLM_PROVIDER=gemini|anthropic|huggingface.

The terminal prints a diagnostic block on every startup showing which keys were found and which provider/model is active — check that first if something's not working.

Known limitations by provider

• Gemini (free tier): rate-limited; some projects get a 0-quota free tier until billing is linked on the Google Cloud project (see README troubleshooting history / ask if you hit a 429 with limit: 0).
• Anthropic: no free tier — pay-as-you-go per token, small one-time trial credit on new accounts.
• Hugging Face (free tier): no billing required, but weaker instruction-following than Claude/Gemini — the "answer only from the document, else say not found" behavior is less reliable with smaller open models. Larger instruct models (like the Llama-3.3-70B default) do best. Cold starts of a few seconds are possible on first request.

Open http://localhost:5000 in your browser.

How it works

1. Upload your PDF — sent to POST /api/extract, text is pulled out with pypdf.

2. Build Knowledge Base — the first ~6,000 characters are sent to Claude (POST /api/build-kb) with instructions to return a short domain label (e.g. "Airline HR Policy").

3. User Query — POST /api/query. If the document is long, TF-IDF cosine similarity picks the most relevant chunks first; Claude is instructed to answer only from that content, and to return a fixed sentinel token if the answer genuinely isn't in the document. The backend converts that sentinel into:

   Dear User, Apologies. The searched content is unavailable in the uploaded document.

4. Refresh icon in the header reloads the page and clears all state.

Deploying publicly (Render)

This project includes a render.yaml blueprint, so deployment is largely one-click:

1. Push the project to a GitHub repo (.env is excluded via .gitignore — never commit it).
2. On render.com, choose New → Blueprint, point it at your repo. Render reads render.yaml automatically and sets up the web service, build command, and start command for you.
3. During setup Render will prompt for the env vars marked sync: false (HF_API_KEY, GOOGLE_API_KEY, ANTHROPIC_API_KEY, LLM_PROVIDER). You only need to fill in the ONE you're actually using — leave the other two blank.
4. Deploy. You'll get a public URL like https://paper-mind-ai.onrender.com.

Notes:

• Render's free tier sleeps after ~15 minutes of inactivity; the first request after that takes 30–50 seconds to wake back up.

• If you'd rather set things up manually instead of via the blueprint: New → Web Service → select repo → Build Command pip install -r requirements.txt → Start Command gunicorn app:app → add your key under the Environment tab.

• This demo keeps everything in memory in the browser tab (document text is round-tripped with each request) rather than a database — fine for single-user local use. For multi-user or production use, add server-side session storage.

• Swap in your own logo at static/ and reference it from index.html if you'd like the exact brand asset instead of the built-in SVG mark.