An LLM That Designs Its Own Data Tools

What it does

This LLM-based system handles end-to-end data tasks for Chinese financial markets—stocks, funds, economic indicators, and financial data. Submit a natural-language request, and it autonomously manages, processes, analyzes, predicts, and visualizes the answer, delivering graphics, tables, and text without manual pipeline wiring. It currently supports GPT-3.5, Azure GPT-3.5, and Qwen-72b-Chat as back-end models.

The interesting bit

Instead of relying on pre-built toolchains, the system acts as both a designer and a dispatcher: it iteratively refines its own interface tools through self-request, then invokes them in parallel or sequence to transform heterogeneous raw data into polished outputs. The LLM is essentially drafting its own API client on the fly to match your intent.

Key highlights

• Autonomous workflow: handles querying, prediction, analysis, and visualization in one pass
• Self-designing tools: the LLM iteratively refines interface tools via “self-request” rather than using a fixed library
• Chinese market focus: connects to stocks, funds, economic, and financial data through Tushare
• Multi-model support: works with OpenAI GPT-3.5, Azure GPT-3.5, and Qwen-72b-Chat
• Hosted demo available on Hugging Face Spaces with a Gradio interface

Caveats

• GPT-3.5’s 4k token limit currently restricts how much data the system can access and process in one go
• Scope is limited to Chinese markets for now; foreign financial data is listed as future work
• Requires valid API keys for both the LLM provider and the Tushare data source

Verdict

Worth a look if you’re building autonomous data agents or need LLM-driven analysis of Chinese financial datasets. Skip it if you need production-grade access to non-Chinese markets or cannot rely on external API tokens.