Course Project: IIIT LLM Course
Transform a short product idea into a comprehensive Product Requirements Document using multi-agent AI with RAG-powered context retrieval.
Idea2PRD is a Streamlit web application that converts a brief product idea into a structured, professional-grade Product Requirements Document (PRD). It uses a multi-agent architecture powered by LLMs (Gemini or OpenAI) where three specialized AI agents collaborate sequentially:
- Business Analyst Agent — Extracts problem context, target users, personas, and pain points
- Product Manager Agent — Generates user stories, functional/non-functional requirements, and MVP scope
- Risk Reviewer Agent — Identifies risks, ambiguities, missing requirements, and clarifying questions
The system supports RAG (Retrieval-Augmented Generation) — users can upload supporting documents (competitor analyses, market research, business rules) that are chunked, embedded, and retrieved to ground the PRD in real context.
| Feature | Description |
|---|---|
| 🤖 Multi-Agent Pipeline | 3 specialized agents (BA, PM, Risk) running sequentially |
| 📄 RAG Support | Upload PDF/TXT documents for context-grounded generation |
| 🔍 Citation Tracking | Shows which parts of the PRD used retrieved document context |
| 📊 Evaluation Framework | Built-in quality checklist and manual evaluation rubric |
| ⬇️ Export | Download the generated PRD as a Markdown file |
| 🧪 Sample Inputs | 3 pre-built product ideas for quick testing |
| 🔑 Multi-LLM Support | Works with Google Gemini, OpenAI, and Groq APIs |
| 🎨 Professional UI | Clean, polished Streamlit interface with custom styling |
User Input (Idea + Users + Domain + Constraints)
│
▼
┌─────────────────────────────────────────┐
│ Streamlit Frontend │
│ (Sidebar: Config, Upload, Samples) │
│ (Main: Input Form, Results Tabs) │
└───────────┬─────────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ RAG Pipeline (rag_utils.py) │
│ PDF/TXT Parsing → Chunking → FAISS │
│ → Similarity Search → Context │
└───────────┬─────────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ Multi-Agent Workflow (agents.py) │
│ │
│ ┌─────────────┐ │
│ │ Business │──→ Problem Statement │
│ │ Analyst │ Personas, Pain Pts │
│ └──────┬──────┘ │
│ ▼ │
│ ┌─────────────┐ │
│ │ Product │──→ User Stories, │
│ │ Manager │ Requirements, MVP │
│ └──────┬──────┘ │
│ ▼ │
│ ┌─────────────┐ │
│ │ Risk │──→ Risks, Gaps, │
│ │ Reviewer │ Clarifying Qs │
│ └──────┬──────┘ │
│ ▼ │
│ ┌─────────────┐ │
│ │ Compiler │──→ Final Merged PRD │
│ └─────────────┘ │
└───────────┬─────────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ LLM (Gemini 2.0 Flash / GPT-4o) │
└─────────────────────────────────────────┘
│
▼
📄 Structured PRD Output
(12 sections, citations, export)
For the detailed architecture, see architecture.md.
- Python 3.10+
- A Google Gemini API key (Get one free) or an OpenAI API key or Grok Api key
# 1. Clone the repository
git clone <your-repo-url>
cd Idea-prd
# 2. Create a virtual environment (recommended)
python -m venv venv
source venv/bin/activate # macOS/Linux
# venv\Scripts\activate # Windows
# 3. Install dependencies
pip install -r requirements.txt
# 4. Run the application
streamlit run app.py- Open the app in your browser (usually
http://localhost:8501) - Enter your Gemini or OpenAI API key in the sidebar
- Select a sample input from the sidebar dropdown (or type your own idea)
- Optionally upload the sample documents from the
sample_data/folder - Click 🚀 Generate PRD
- Explore the results across all 4 tabs
Idea-prd/
├── app.py # Main Streamlit application
├── agents.py # Multi-agent workflow (BA, PM, Risk)
├── prompts.py # All structured system prompts
├── rag_utils.py # PDF/TXT parsing, chunking, FAISS, retrieval
├── config.py # App configuration and constants
├── requirements.txt # Python dependencies
├── README.md # This file
├── architecture.md # Detailed architecture document
├── sample_data/
│ ├── competitor_analysis.txt
│ ├── market_research.txt
│ └── business_rules.txt
└── .streamlit/
└── config.toml # Streamlit theme configuration
| Screen | Description |
|---|---|
 |
Product idea input form |
 |
Generated PRD output |
 |
Retrieved document chunks |
 |
Individual agent results |
 |
Quality evaluation framework |
Follow this script to demonstrate the full capabilities:
- Start the app:
streamlit run app.py - Configure: Enter your Gemini API key in the sidebar
- Upload documents: Upload all 3 files from
sample_data/ - Load sample input: Select "🎓 AI Study Planner" from the sidebar
- Generate: Click "🚀 Generate PRD"
- Review tabs:
- Final PRD: Show the complete, merged document
- Retrieved Context: Highlight the chunks retrieved from uploaded docs
- Agent Outputs: Expand each agent to show their individual analysis
- Evaluation: Walk through the quality checklist
- Export: Download the PRD as markdown
- Compare: Run again without uploaded documents — note the "[Assumption]" markers
- Structured system prompts with explicit output format requirements
- Role-based personas for each agent (Business Analyst, Product Manager, Risk Reviewer)
- Guardrails against hallucination: agents must cite sources or mark assumptions
- Consistent markdown formatting enforced across all agent outputs
- Documents are parsed (PDF via PyPDF2, TXT via UTF-8 decode)
- Text is chunked using LangChain's
RecursiveCharacterTextSplitter(1000 chars, 200 overlap) - Chunks are embedded using Google's
embedding-001or OpenAI'stext-embedding-3-small - FAISS index enables fast similarity search
- Top-5 chunks are retrieved per agent query and injected into prompts
- Sequential pipeline: BA → PM → Risk → Compiler
- Each agent receives prior agents' outputs as context
- Each agent gets independently retrieved RAG context
- Final compiler merges all outputs into a cohesive document
- Token limits: Very long documents may be truncated during chunking
- Single LLM: All agents use the same model (no specialized models per agent)
- No memory: Agents don't retain context between sessions
- English only: Prompts and output are optimized for English
- No iterative refinement: No feedback loop for PRD improvement
- RAG quality: Effectiveness depends on document quality and relevance to the idea
- Iterative refinement with user feedback
- Support for DOCX, HTML, and URL ingestion
- Agent-to-agent debate for conflict resolution
- Semantic caching for repeated queries
- PRD comparison mode (with vs. without RAG)
- PDF export with professional formatting
- Industry-specific prompt templates
- Multi-language support
This project is created for academic purposes as part of the IIIT LLM Course.
Built with ❤️ by Parmod Kumar using LangChain, FAISS, Streamlit, and Gemini/OpenAI/groq