Challenges in Implementing AI Assistant for Data Querying and Chart Generation with EdgeQL

[Jady891213]

We're building a system using GEL and want to implement an AI assistant for data querying and chart generation, but are facing several challenges:

LLM Limitations with EdgeQL

General-purpose LLMs are only familiar with SQL, and their training datasets primarily contain SQL-related materials. We attempted to generate EdgeQL using knowledge bases, but the accuracy is unacceptably low – especially for grouping syntax (GROUP, count(.elements), etc.), which fails completely.

Ineffective Fine-tuning

We compiled a small set of EdgeQL training examples for LLM fine-tuning, but the results remain unsatisfactory.

PostgreSQL Bridge Limitations

When utilizing GEL's PostgreSQL compatibility layer:

The generated SQL cannot leverage EdgeDB's unique features:
Computed properties/links
Global variables
Type system dependencies
Results in capability parity loss compared to native EdgeQL

Current AI Extension Limitations

GEL's existing AI extensions appear limited to:

Vector data operations
Object-level semantic search ("Find products similar to X")
While lacking support for free-form queries like:
"Show last month's revenue and costs per channel"
which require dynamic EdgeQL generation.
Has the GEL team considered solutions addressing these challenges? We would appreciate guidance on either:

Improving EdgeQL generation accuracy with LLMs, or
Expanding GEL's AI-native query capabilities.