The Project
At Input Output Global (IOG), there are many documentation sites covering multiple projects and knowledge sources. A common question was, “Which doc site should I look at to find what I’m looking for, and how do I find it?” To address that reader pain point, I designed and built the Essential Cardano AI Assistant — a RAG chatbot that could synthesize answers, with citations, from 2,572 files across eight documentation sites:| Source | URL | Files |
|---|---|---|
| Essential Cardano | essentialcardano.io | 979 |
| IOG Research | iog.io/en/research | 625 |
| Intersect Knowledge Base | docs.intersectmbo.org | 354 |
| Developer Resources | developers.cardano.org | 251 |
| Intersect News | intersectmbo.org/news | 100 |
| IOG News | iog.io/news | 95 |
| Plinth User Guide | plutus.cardano.intersectmbo.org | 85 |
| Cardano Docs | docs.cardano.org | 83 |
What I Delivered
1. Full-Stack RAG Pipeline
I built a complete retrieval-augmented generation system from the ground up:- Indexed 2,572 documents from 8 sources across the Cardano ecosystem
- Semantic search with citation generation — every answer links back to the source document
- Next.js with the Vercel AI SDK and assistant-ui component library for the chat interface, styled with Tailwind
- Deployed on Railway with a live demo interface
Live demo — knowledge base current as of December 2025.
2. Citation-First Design
Every response links back to source documents — users can verify claims with one click. To make citations meaningful, I built date awareness into the knowledge base:- Extracted and normalized publish dates across 2,500+ source files
- Prioritized recent content in retrieval to avoid stale answers
- Consistent date citation format so users know how current the information is
- 20+ system prompt iterations to enforce citation standards and IOG’s editorial style
3. Documentation
Created comprehensive documentation for the system:- User guide for developers
- System architecture overview
- Prompt engineering decisions and rationale
- Limitations documentation (what the assistant will and won’t answer)
2,572 Documents
Indexed across 8 sources: Essential Cardano, IOG Research, Intersect Governance, Developer Portal, and more.
Date-Aware Pipeline
Publish dates extracted, recent content prioritized, consistent citation format.
Citation Generation
Every answer links back to source documents. Users can verify and dive deeper.
20+ Prompt Iterations
System prompt refined to enforce citation standards, date formatting, and IOG’s editorial style.
Tech Stack
Python · Vector Databases · RAG Pipelines · Prompt Engineering · Railway · Anthropic Claude