🤖 AI Agent — Chat with your PDFs

A conversational AI agent powered by OpenAI's GPT-4o that can read your PDF documents and answer questions about them using RAG (Retrieval-Augmented Generation). Also supports web search and persistent memory across sessions.

---

🌐 Live Demo

Try it here: pdfragagent.streamlit.app

Bring your own key: the live demo doesn't ship an API key — paste your own OpenAI key in the sidebar (it's used only for your session and never stored).

Upload a PDF in the sidebar, then ask questions about it — or ask anything and let the agent search the web.

Deploying for the first time? See DEPLOY.md for a 5-minute Streamlit Cloud walkthrough.

📸 Screenshot

Placeholder — replace docs/screenshot.png with an actual screenshot of your deployed app.

---

✨ Features

• 📄 PDF Question Answering (RAG) — Upload PDFs and ask natural language questions. The agent finds the most relevant sections and generates accurate answers with source citations.
• 🔍 Web Search — Searches the web via DuckDuckGo when your documents don't have the answer. No extra API key required.
• 🧠 Persistent Memory — Conversation history is saved to disk. Restart the app and it picks up right where you left off.
• ⚡ Smart Caching — PDF embeddings are cached locally. Load a file once and it's instantly available on every restart.
• 🛠️ Function Calling — The agent automatically decides whether to search your documents, the web, or just reply from memory. You don't have to tell it what to do.

---

🏗️ Architecture

User Question
     │
     ▼
┌──────────┐
│  GPT-4o  │ ──── decides which tool to use
└──────────┘
     │
     ├──▶ search_documents  ──▶  Embedding similarity search over PDF chunks
     │
     └──▶ web_search        ──▶  DuckDuckGo real-time results
     │
     ▼
Final Answer (with sources)

RAG Pipeline:

1. PDFs are extracted → split into overlapping chunks (~800 chars)
2. Each chunk is embedded using text-embedding-3-small
3. On query, the user's question is embedded and compared via cosine similarity
4. Top 5 matching chunks are sent to GPT-4o as context
5. The model generates an answer citing the source PDF

---

🚀 Quick Start

1. Clone the repo

git clone https://github.com/amanparganiha/PdfRagAgent.git
cd PdfRagAgent

2. Install dependencies

pip install -r requirements.txt

3. Set up your API key

cp .env.example .env

Open .env and paste your OpenAI API key:

OPENAI_API_KEY=sk-proj-your-actual-key-here
MODEL=gpt-4o

Get your key at platform.openai.com/api-keys

4. Add your PDFs

Drop PDF files into the pdfs/ folder (created automatically on first run):

your-project/
├── chatbot.py
├── pdfs/
│   ├── research_paper.pdf
│   └── company_report.pdf
└── ...

5. Run

Web app (recommended):

streamlit run app.py

Opens at http://localhost:8501 — upload PDFs and chat in the browser. This is also what gets deployed (see DEPLOY.md).

Terminal version:

python chatbot.py

---

💬 Usage

Chat normally

🧑 You: What is the main conclusion of the research paper?
🤖 Agent: According to research_paper.pdf, the main conclusion is...

Ask about the web

🧑 You: What's the latest news about OpenAI?
🤖 Agent: 🔍 Searching web...

Load a PDF mid-conversation

🧑 You: /load C:\Users\me\Documents\notes.pdf
📄 Processing: notes.pdf
   ✂️ Split into 24 chunks
   📊 Embedded 24/24 chunks
   ✅ Done!

Slash Commands

Command	Description
/load <path>	Load a PDF file into the agent
/docs	List all loaded documents and chunk counts
/clear	Reset conversation memory
/history	View last 10 messages
quit	Exit the agent

---

⚙️ Configuration

All settings are in the top section of chatbot.py:

Setting	Default	Description
EMBEDDING_MODEL	text-embedding-3-small	OpenAI embedding model
CHUNK_SIZE	800	Characters per text chunk
CHUNK_OVERLAP	200	Overlap between chunks for context continuity
TOP_K	5	Number of chunks retrieved per query
MAX_HISTORY	50	Max messages kept in conversation memory

You can also change MODEL in your .env file to use gpt-4o-mini or gpt-3.5-turbo for cheaper usage.

---

📁 Project Structure

.
├── app.py              # Streamlit web UI (deployed version)
├── chatbot.py          # Terminal/CLI version
├── requirements.txt    # Python dependencies
├── DEPLOY.md           # Deployment guide (Streamlit Cloud, etc.)
├── .env.example        # Template for API keys (CLI / local)
├── .env                # Your actual keys (git-ignored)
├── .gitignore          # Keeps secrets & caches out of git
├── .streamlit/
│   └── secrets.toml.example   # Key template for Streamlit (git-ignored when real)
├── docs/
│   └── screenshot.png  # App screenshot for the README
├── pdfs/               # Drop your PDF files here, CLI only (git-ignored)
└── memory/             # CLI persistence (git-ignored)
    ├── chat_history.json   # Conversation memory
    └── embeddings/         # Cached PDF embeddings

Note: The web app (app.py) keeps chat and embeddings in the browser session rather than the memory/ folder, since cloud hosts have ephemeral disks. The pdfs/ and memory/ folders are used by the CLI version.

---

💰 Cost Estimate

Action	Model	Approx. Cost
Embed a 50-page PDF	text-embedding-3-small	~$0.002
One chat message	gpt-4o	~$0.01–0.03
One chat message	gpt-4o-mini	~$0.001
Web search	DuckDuckGo	Free

Embeddings are cached, so you only pay once per PDF.

---

🛡️ Security

• API keys are stored in .env and never committed to git
• .gitignore excludes .env, memory/, and pdfs/
• No data is sent anywhere except OpenAI's API

⚠️ Never hardcode your API key in the source code. If you accidentally push a key to GitHub, revoke it immediately at platform.openai.com/api-keys.

---

🤝 Contributing

Pull requests are welcome! Some ideas for improvements:

• Support for more file types (Word, TXT, CSV)
• Streaming responses
• Web UI with Streamlit or Gradio
• Multiple conversation threads
• Summarize entire documents
• Token usage tracking

---

📄 License

MIT License — use it however you want.