← Head back to all of our AI Engineer World's Fair recaps
Tanmai Gopal @tanmaigo / Hasura
Watch it on YouTube | AI.Engineer Talk Details
Tanmai from Hasura gave a talk about connecting real-time data to AI systems.
He focused on the challenges of making AI models work effectively with live data sources and introduced Hasura's solution for this problem.
Key Points
- Problem: Current AI models struggle to interact intelligently with real-world data sources like calendars, CRM systems, and product databases.
- Solution: Make live data and business logic available as a tool for language models (LLMs).
- Core Ideas:
- Unified Query Language
- Object Model for Authorization
- LLM-driven Data Access Planning
Demo
Tanmai demonstrated a system where an AI could interact with a fictional Blockbuster movie rental database.
The AI was able to write a personalized email to a top customer, mentioning their recent movie rentals, by accessing multiple data sources.
Check out the Hasura Pacha repo over on GitHub
Technical Insights
- Unified Query Language:
- Use SQL-like syntax for all data interactions (structured, unstructured, APIs)
- LLMs understand SQL semantics better than custom API parameters
- Authorization Model:
- Data schema and session properties determine access rules
- Consistent authorization across all access methods
- LLM-driven Data Access:
- Instead of hard-coding data retrieval, ask the LLM to write Python code to fetch required data
- This approach leverages the LLM's code generation abilities more effectively than direct reasoning
This talk offered an interesting perspective on bridging the gap between AI models and live data sources. The approach of using a unified query language and letting AI generate data access code could be a game-changer for many applications.
It's definitely an area to watch as AI systems become more integrated with real-world data and business processes.