How to Build an Agent That Better Understands Data Tables

In the data era, every table may hide huge business value. However, how to quickly and accurately gain insights from massive tables has always been a difficult problem troubling many enterprises. AskTable's mission is to enable everyone to easily and enjoyably gain insights from data.

How do we do it? The answer is not simply connecting a Large Language Model (LLM) directly to a database. Instead, we built a rigorous, reliable and efficient Agent system. The core idea is: let AI do what it's good at, and place it in a controllable, verifiable "cage."

Avoiding Hallucinations: Generate Code, Not Content

At AskTable, we firmly believe "accuracy is everything." LLMs are essentially language models; letting them directly generate data analysis conclusions easily produces "speaking nonsense with a straight face" (hallucinations).

Our solution is: don't let LLM directly touch, calculate, or generate final data; instead, guide it to generate executable code (like SQL or Python).

The process is roughly as follows:

1. •User asks a question
2. •→ LLM understands and generates query or computing code
3. •→ Code is executed in a secure sandbox environment
4. •→ Execution results are rendered into charts and text
5. •→ Presented to the user

This approach has many benefits:

• •Verifiability: All conclusions have traceable code, ensuring authenticity and accuracy of results.
• •Stability: Fully leverages the high performance and stability accumulated over years by databases (like MySQL, PostgreSQL, ClickHouse).
• •Security: Combines AI's creativity with engineering rigor, effectively avoiding hallucinations.

"Clear, accurate, and complete context is the best means to avoid hallucinations." Through this model, we provide LLM with the most effective context—structured code and data.

Dual Engine: Handling Different Analysis Scenarios

To meet the needs of different users, we designed two collaborative modes:

• •

  Instant Q&A Mode For business personnel's daily high-frequency queries. Using "fixed but agile Workflow," common analysis processes are templated, achieving low-cost, instant response (from minute-level to second-level). Like an ever-tireless data assistant.

• •

  Explora Mode (Exploratory Analysis) (released in September) For data analysts' in-depth exploration scenarios. Enables "flexible but controllable Agent" for complex business analysis. This is backed by a sophisticated multi-agent collaboration system.

Agentic RAG + CodeAct: The Agent's Brain and Hands

Our Agent is not a single "large and comprehensive" model, but a team with clear division of labor:

• •

  The "Hands" for Reasoning - CodeAct Compared to ReAct (Think-Act-Observe) popular in the industry, CodeAct is more suitable for precise data analysis scenarios. It uses executable code as the primary "action," ensuring every step is precisely verifiable.

• •

  The "Brain" for Collaboration - Agentic RAG Team

  • •Routing Agent: Like an intelligent dispatcher, judges which tool/database the user's question should be handed to.
  • •Query Planning Agent: Acts as a project manager, breaking complex analysis tasks into executable sub-tasks.
  • •Synthesis/Post-processing Agent: The final "report writer," integrating all step results into clear and coherent conclusions (chart reports).

Three-Layer Memory System: Making the Agent "Photographic Memory" and Continuously Evolving

No one likes repeating themselves, so how do we give AI "memory"?

We designed a three-layer memory system:

1. •In-Context Memory (RAM): Maintains short-term memory within a single conversation, understanding context.
2. •Long-term Semantic Memory (Hard Drive): Let AI "understand you." Not only stores key business knowledge across conversations, but also automatically distills personalized preferences. Just say it once, AskTable will remember.
3. •Agentic, Evolutionary Memory (Learning Brain): Let the memory system itself have learning and evolutionary capabilities, achieving true "personalization" (future direction).

Memory is one of the core capabilities of future AI, a difficult but right thing to do. The company name "Memory Future" embodies this belief.

Dual Evaluation System: Ensuring Model Suitability and Solution Feasibility

To be responsible to customers, we established a strict dual test evaluation system:

• •Model Selection Evaluation: Most customers want private model deployment. This system helps select models suitable for AskTable (including "Basic" and "Standard" two solutions), filtering from latency, cost and effectiveness three dimensions, ensuring that even with limited budgets, customers can enjoy AI convenience.
• •Solution Feasibility Evaluation: End-to-end testing of actual user scenario questions, honestly and clearly displaying capability boundaries, ensuring what AskTable can and cannot solve for customers.

Conclusion

From the underlying database, vector storage, keyword search, to the upper agent architecture and dual engine, and our adherence to the core philosophy of "putting AI in a cage," every step of AskTable aims to build a reliable, powerful and easy-to-use table agent.

We don't pursue uncontrollable "magic," but are committed to safely and pragmatically empowering everyone who needs to converse with data with AI's powerful capabilities through excellent engineering practices. We believe this is the correct path to true data democratization.