An AI-powered internal tool that answers questions about company documents, summarizes meeting notes, and extracts actionable tasks with owners and priorities.
Built as a RAG (Retrieval Augmented Generation) system using LangChain, ChromaDB, and FastAPI — with a clean browser frontend.
Document Q&A Upload any PDF and ask questions about it in natural language. The system retrieves only the relevant sections and answers based strictly on document content — it won't hallucinate information that isn't there.
Meeting Summarizer Paste raw meeting notes or a transcript in any format. Get back a structured summary with key decisions, action items, owners, and next steps.
Task Extraction Toggle on task extraction to get a structured list of action items with assigned owners, due dates, and priorities — formatted as JSON ready to feed into any calendar or project management tool.
| Component | Technology | Purpose |
|---|---|---|
| API Framework | FastAPI | REST endpoints, automatic Swagger docs |
| RAG Framework | LangChain | Pipeline orchestration |
| Vector Database | ChromaDB | Local semantic search |
| Embeddings | OpenAI text-embedding-ada-002 | Convert text to vectors |
| LLM | GPT-3.5-turbo | Answer and summary generation |
| Frontend | HTML + CSS + JavaScript | Browser UI, served by FastAPI |
| Runtime | Python 3.11 | — |
PDF Upload
↓
PyPDFLoader → text extraction + image-PDF detection
↓
RecursiveCharacterTextSplitter → 500 char chunks, 50 char overlap
↓
OpenAI Embeddings → vectors
↓
ChromaDB → vector storage (collection cleared via delete_collection() on each upload)
─── Query time ───
User question → vector → ChromaDB similarity search → top 3 chunks
↓
Chunks + question → GPT-3.5-turbo (temperature=0)
↓
Grounded answer
| Method | Endpoint | Description |
|---|---|---|
| GET | / |
Health check + database status |
| POST | /upload |
Upload PDF to knowledge base |
| POST | /ask |
Ask a question about uploaded document |
| POST | /summarize |
Summarize meeting notes |
| POST | /extract-tasks |
Extract structured task list from meeting notes |
| GET | /ui |
Browser frontend |
Chunk size: 500 characters with 50 character overlap Tested sizes from 50 to 1000. Smaller chunks improve retrieval precision but lose surrounding context, resulting in incomplete answers. 500 with overlap balances precision and completeness.
Temperature: 0 Set to 0 for deterministic, factual responses. Higher temperature increases creativity but also hallucination risk — wrong for a document Q&A use case.
Strict prompt constraints The prompt explicitly instructs the model to answer only from provided context. Tested with out-of-scope questions — system correctly returns "not enough information" rather than falling back to training data.
Replace-on-upload strategy
Each upload clears the existing ChromaDB collection via delete_collection()
and rebuilds it from the new document. This avoids data consistency issues
where multiple versions of the same document would cause conflicting answers.
Using the ChromaDB API rather than deleting the folder keeps the SQLite
connection stable within the same process.
Retry with exponential backoff All LLM calls are wrapped in retry logic with exponential backoff (1s, 2s, 4s). Handles transient API failures gracefully without surfacing errors to the user unnecessarily.
Input sanitization at API boundary All text inputs are sanitized to remove invisible control characters before processing. This makes the API robust to real-world copy-pasted text from chat tools and meeting platforms.
- File type validation before processing
- Image-based PDF detection — rejects PDFs with no readable text
- Input sanitization on all text endpoints
- Empty input validation on all endpoints
- Try/except on all LLM and database calls with clean HTTP error responses
- Finally blocks for temp file cleanup on upload
- Retry logic with exponential backoff on all LLM calls
Requirements: Python 3.11, OpenAI API key
# Clone and enter directory
git clone https://github.com/[yourusername]/document-qa
cd document-qa
# Create virtual environment
python3.11 -m venv venv
source venv/bin/activate
# Install dependencies
pip install -r requirements.txt
# Add your OpenAI API key
echo "OPENAI_API_KEY=your-key-here" > .env
# Run
uvicorn main:app --reload
# Run tests
pytest tests/ -vOpen http://localhost:8000/ui for the browser interface.
Open http://localhost:8000/docs for the Swagger API explorer.
- RAG consistency depends on both LLM temperature AND retrieval stability — specific factual questions are more stable than broad ones because fewer chunks compete for relevance
- Chunk size affects answer completeness not retrieval accuracy — the vector search finds the right location regardless, but small chunks don't carry enough surrounding context
- Prompt constraints are the primary defense against hallucination — "answer only from context" is more reliable than relying on the model's own judgment
- Data consistency in multi-document RAG requires careful design — conflicting versions of the same document produce inconsistent answers, which led to the replace-on-upload decision
- Real-world text input from chat tools and meeting platforms contains invisible control characters that break JSON parsing — input sanitization at the API boundary is essential for production robustness
- Structured JSON output from LLM calls requires defensive parsing — models don't always return clean JSON even when instructed to
- Google Calendar integration via MCP for automatic task scheduling
- Multi-document support with source tracking and versioning
- Evaluation framework using RAGAs for automated quality measurement
- Rate limiting and token usage monitoring
- Q&A history per document with automatic reset on new upload
- Make LLM calls async to avoid blocking the event loop under concurrent load