APItome

Inspiration

I just learnt about RAG and Generative AI. I wanted to apply my learning to create a tool that solves a real problem I (and many other devs) face - integrating various third-party APIs seamlessly. Developers have to constantly juggle through multiple API docs, while AI agents often lack up-to-date context on the latest APIs.

What it does

APItome is a GenAI + RAG powered chatbot that allows users to:

Submit multiple public API documentation URLs
Crawl and index the documentation into a unified knowledge base
Ask cross-API questions conversationally

For example:

“Get weather for coordinates from Google Maps location search.”
“How do I authenticate Spotify and YouTube APIs together?”
“Compare rate limits between OpenWeather and WeatherAPI.”

Instead of jumping across tabs, developers get context-aware, grounded answers from all indexed APIs in one place.

How I built it

I built it as a Retrieval Augmented Generation (RAG) application that transforms API documentation into a unified conversational assistant. My role was primarily designing the architecture, features and the implementation plan. I decided to go with a micro-services based architecture and made it into 3 services -

Parser Service: Handles web crawling and documentation ingestion. Accepts API doc URLs, extracts structured content (endpoints, parameters, examples), cleans and chunks the data, and stores the processed embeddings in ChromaDB for semantic retrieval.
RAG Service: Manages the core retrieval and generation pipeline. Embeds user queries, retrieves relevant chunks across all indexed APIs from ChromaDB, constructs context-aware prompts, and interfaces with Groq's Llama 3.1 to generate grounded, cross-API responses with streaming support.
Gateway Service: Acts as the unified FastAPI entry point for the frontend. Handles user authentication, per-user rate limiting, conversational context management, and routes requests to the appropriate downstream service.

I also decided to have a separate frontend application. The code for all of these were done by AI agents like Claude, using tools such as Github Copilot. This helped me built the entire working product in a few days.

Challenges I ran into

Parsing content from documentation websites was a major challenge as websites don't have a standard format
Managing a huge codebase primarily coded using AI required multiple AI agents, context + right prompting was a task at times
Fine-tuning the RAG and the LLM to ensure it matches my expectations, and also doesn't answer anything out of context