📄 DocChat - RAG Powered Document Q&A

Upload any PDF. Ask questions in plain English. Get answers grounded in the document with exact page citations.

---

🖥️ UI Preview

In the screenshot above, a resume PDF was uploaded and indexed into 6 chunks inside ChromaDB. The question "Tell me about the candidate's focused domain and relevant experience?" was asked — DocChat retrieved the 4 most relevant chunks via cosine similarity and passed them as context to Llama 3.3 70B, which generated a structured, grounded answer with bullet points pulled directly from the document. No hallucination — every fact comes from the PDF.

👉 Try the live app — bring your own OpenRouter API key (free tier available)

---

🧠 How It Works

┌─────────────────────────────────────────────────────────┐
│                      PDF Upload                          │
└───────────────────────┬─────────────────────────────────┘
                        │
                        ▼
              pypdf — extract text per page
                        │
                        ▼
         RecursiveCharacterTextSplitter
          chunk_size=1000, overlap=200
                        │
                        ▼
        text-embedding-3-small (OpenRouter)
          → dense vector per chunk
                        │
                        ▼
                   ChromaDB
            cosine similarity index
                        │
                  (on each query)
                        │
                        ▼
         top-4 chunks retrieved by similarity
                        │
                        ▼
         Llama 3.3 70B via OpenRouter
      grounded answer + page citations

---

🛠️ Tech Stack

Layer	Tool	Purpose
UI	Streamlit	Chat interface + PDF upload
PDF Parsing	pypdf	Extract text per page with metadata
Chunking	LangChain RecursiveCharacterTextSplitter	1000-char chunks, 200-char overlap
Embeddings	text-embedding-3-small via OpenRouter	Dense vector representation
Vector Store	ChromaDB	Cosine similarity retrieval
LLM	Llama 3.3 70B via OpenRouter	Answer generation with grounding

---

🚀 Run Locally

1. Clone & set up environment

git clone https://github.com/dabhiram13/DocChat.git
cd DocChat
python3 -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

2. Add your API key

cp .env.example .env
# Paste your OpenRouter key into .env

Get a free key at openrouter.ai/keys.

3. Run

streamlit run app.py

Opens at http://localhost:8501

---

📖 Usage

1. Enter your OpenRouter API key in the sidebar
2. Upload a PDF (up to 200 MB)
3. Wait a few seconds — the document is parsed, chunked, embedded, and indexed into ChromaDB
4. Ask any question about the document in the chat
5. Expand 📚 Source passages under each answer to see the exact chunks ChromaDB retrieved and the page numbers they came from

---

📁 Project Structure

DocChat/
├── app.py              # Streamlit UI — sidebar, upload, chat, citations
├── rag_pipeline.py     # RAG core — parse → chunk → embed → store → retrieve → generate
├── requirements.txt    # Pinned dependencies
├── runtime.txt         # Python 3.11 for Streamlit Cloud
├── .env.example        # API key template
├── docs/
│   └── screenshot.png  # UI preview (used in README)
└── .gitignore

---

💡 Key Design Decisions

ChromaDB for vector storage — cosine similarity retrieval gives semantically relevant chunks rather than keyword matches. Each session creates a fresh in-memory collection that's populated when the PDF is uploaded.

Chunk overlap (200 chars) — prevents answers from being cut off at chunk boundaries when context spans two adjacent chunks.

Page-level citations — every answer surfaces the exact page number and text excerpt it was drawn from, so responses are fully verifiable against the source document.

OpenRouter for both LLM and embeddings — single API key, single provider, consistent latency. text-embedding-3-small is fast and cost-efficient; Llama 3.3 70B gives strong reading comprehension for document Q&A.

---

🔑 API Key

This app uses the OpenRouter API for both embeddings and LLM inference.

• Sign up at openrouter.ai — free credits on registration
• Models used: openai/text-embedding-3-small + meta-llama/llama-3.3-70b-instruct
• No data is stored — documents are processed in memory and discarded when the session ends