Potential — AI Search Engine for Abu Dhabi Open Data
A conversational search layer built on top of the Abu Dhabi Open Data Platform. Ask in plain language, get matched datasets, inspect the data — no domain vocabulary required.
Watch it in action
This demo walkthrough shows how a natural language question becomes a retrieved dataset and a chart — in a single conversation turn.
The Potential Pipeline
How it works
A three-stage retrieval-augmented pipeline built on Azure
Query → Keywords
The user types a plain-language request. A first GPT call extracts the machine-readable search string, handling abbreviations, ambiguities, and cross-domain terminology.
Keywords → Datasets
Azure AI Search runs full-text retrieval against the indexed Abu Dhabi Open Data catalog and returns the top-5 most relevant dataset records and their metadata.
Datasets → Answer
A second GPT call synthesizes the retrieved metadata into a human-readable response and highlights the specific dataset identifiers ready for visualization or download.
What it can do
Plain-Language Discovery
No need to know exact dataset names or tags. The system maps natural questions to the closest matching open datasets in the catalog.
Zero-Storage Indexing
Datasets stay on the Abu Dhabi Open Data Platform. The system queries them live through the public API — nothing is copied or cached locally.
In-Browser Visualization
Once a dataset is found, users can render it as a table, bar chart, or line chart directly in the interface without exporting to another tool.
Document-Driven Queries
Upload a PDF or image. Azure Form Recognizer extracts its text and uses it as query context — useful for policy documents or reports.
Multi-Turn Context
Up to six conversation turns are preserved, so follow-up refinements narrow results naturally without restarting from scratch.
Scalable API-First Design
The architecture is not specific to Abu Dhabi. Any open-data platform with a public API can be indexed and queried with minimal reconfiguration.
Tech stack
Frontend
AI & NLP
Data Layer
Charts
Built under pressure
Potential was conceived and shipped in 48 hours during the Abu Dhabi Spark AI Hackathon — a competition run in partnership with government entities. The team of four ranked in the top 10 among more than 26 competing teams. The challenge: make open government data genuinely accessible to non-technical users. The bet was that a thin AI layer on top of existing public APIs could eliminate the steep learning curve of structured data search.