API Documentation QA Agent

A lightweight AI agent that answers natural language questions about APIs using their official documentation. It leverages vector embeddings and Retrieval-Augmented Generation (RAG) for accurate, context-aware responses.

---

🚀 Features

• 📄 Ingests HTML, Markdown, or text-based API docs
• ✂️ Chunks and embeds docs using semantic models
• 🧠 Stores embeddings in LanceDB for fast retrieval
• 🔍 Retrieves relevant context for user queries
• 💬 Uses Groq LLM (LLaMA 3) for natural language answers
• 🧾 Streamlit-based interactive interface
• 🔁 Maintains basic conversation history for follow-ups

---

🧱 Tech Stack

Component	Technology	Overview
Language	Python	Widely used for AI/ML and rapid prototyping; strong library ecosystem
LLM	Groq (LLaMA 3.3 70B Versatile)	This open-source free LLM model provides high-performance inference for complex queries with minimal latency
Agent Framework	Phidata	Lightweight framework that simplifies building agent-based workflows
Vector DB	LanceDB	Chosen for its performance and simplicity; ideal for lightweight, local deployments
Embeddings	SentenceTransformer (all-MiniLM-L6-v2)	Offers a strong balance between speed and accuracy for generating semantic embeddings
Scraping	Unstructured Library	Parses diverse content types like HTML and Markdown into structured text
Chunking	LangChain Text Splitter	Efficiently handles large documents for optimal embedding and retrieval
UI	Streamlit	Enables fast development of clean, interactive UIs without frontend expertise
History Storage	JSON	Lightweight and simple method to manage short-term conversational state

---

🛠️ Installation

1. Clone the Repository

git clone https://github.com/iffatul-anon/API-Documentation-QA-Agent
cd API-Documentation-QA-Agent

2. Setup Virtual Environment

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

3. Configure Environment

Create a .env file:

GROQ_API_KEY=your_groq_api_key

4. Run the Application

streamlit run main.py

---

📋 System Requirements

• Python 3.8 or higher
• 1GB free disk space (for vector database storage)
• Active internet connection (for LLM API access)
• Groq API account and key

🔧 Dependencies

Core dependencies:

• streamlit: Web interface
• phidata: LLM agent framework
• sentence-transformers: Document embedding
• lancedb: Vector database
• langchain: Text chunking
• unstructured: Document parsing
• python-dotenv: Environment management

See requirements.txt for complete list of dependencies.

---

💡 How It Works

Step 1: Data Preparation

• Load and parse documentation files
• Chunk them into meaningful segments (e.g., sections, paragraphs)

Step 2: Embedding & Storage

• Generate vector embeddings for each chunk using embedding model
• Store chunks and embeddings in the vector database

Step 3: Q&A Loop

• Accept natural language question from user
• Embed the question and perform semantic search in the vector DB
• Retrieve top-k relevant documentation chunks
• Construct a prompt with user question + retrieved context + conversation history
• Send prompt to LLM and return answer

---

💡 Usage

Load Documentation

1. Open the app in your browser.
2. Enter an API documentation URL in the sidebar.
3. Click Load Docs to parse and embed content.

Ask Questions

Examples:

• "How do I authenticate with the API?"
• "List available endpoints."
• "What parameters does /users accept?"

Manage Docs

• Use "Clear Chat History" to reset coversation history.

---

📁 Project Structure

api-docs-qa-agent/
├── data/                  # Vector DB & chat history
├── scripts/               # Core logic: ingestion, embedding, querying
│   ├── chunk_docs.py
│   ├── embed_retrive_docs.py
│   ├── load_docs.py
│   ├── query_agent.py
│   └── conversation_load_and_save.py
├── main.py                # Streamlit UI entry point
├── requirements.txt
└── README.md

---

⚠️ Limitations

• Best with structured HTML docs (PDFs not supported)
• Conversation history limited (last 5 messages due to prompt size limitations in open-source LLM models)
• Initial load time varies by doc size
• Complex queries may need step-by-step interaction

---

🧰 Troubleshooting

Issue	Fix
App won't start	Check Python version, activate venv, verify dependencies
Doc load fails	Confirm URL accessibility, internet, retry with another URL
No agent response	Verify API key, internet, and doc loading
Poor answers	Rephrase question or load additional docs

---

📬 Contact

• Name: Md. Iffatul Islam Anon
• Email: anon35-1065@diu.edu.bd
• Email: iffatulislamanon@gmail.com
• Linkedin: iffatul-anon

---