Farewell to ChatBI, Reconstructing Data Thinking: The Paradigm Reconstruction of Data Analysis Work in the AI Era and AskTable's Design Philosophy

Before entering the AI field, I was always engaged in model and algorithm development, holding a very traditional Computer Science (CS) perspective. With the prevalence of large models, we observed that AI has made the fastest progress in programming (like Cursor, Claude code), because in this closed loop, we develop, use, test and optimize data ourselves. At that time, we naturally had an idea: since AI performs so amazingly in the code field, in the big data analysis field, wouldn't AI be "dimensionality reduction combat"?

But in the actual implementation process, we found this idea too idealistic.

---

1. Disillusionment and Reflection: Why Isn't "Conversation" the Ultimate Form of Data Analysis?

In the past year, many Text-to-SQL or Chat-to-Data products have emerged in the industry. The technical logic seems perfect: user asks question -> AI translates -> database executes -> returns result. But in actual implementation, the AskTable team identified two core pain points, which prompted a fundamental adjustment of the technical route.

1. Implementation Dilemma: Entropy Increase in Organizational Collaboration

Ideally, AI is a bridge connecting business personnel and databases. But in reality, implementing an AI data project requires coordinating business parties (requirements), data teams (data standards), Ops teams (permissions) and bosses (results). Technical support teams are often drowned in cross-departmental communication noise. AI hasn't automatically eliminated the barriers of organizational structure; instead, due to high requirements for data accuracy, it has increased communication costs caused by missing context.

2. The Interaction Paradox: Linear Conversation vs. Divergent Thinking

This is a deeper technical philosophy issue.

• •Human thinking mode: Data analysis is a "modeling" process. Starting from fuzzy mental representations, through divergence, exploration, trial and error, and finally converging to a verifiable conclusion (like SQL or charts). This is a non-linear, incremental process.
• •Chat's interaction mode: Traditional dialog boxes are linear (Q&A). Users must complete "dimensionality reduction" and "cleaning" in their minds before asking questions, compressing complex business demands into one sentence of Prompt.

Conclusion: Conversation (Chat) is not how humans think about data. Forcing linear conversation to carry divergent analytical thinking adds to users' cognitive load rather than reducing it.

---

2. Architecture Reconstruction: From "Conversation Flow" to "What You Think is What You Get" Canvas

Through observing user feedback, we realized: conversation is not how humans think about data.

1. Business Queries are a Non-Linear Process

Business queries usually follow a process of "receiving requirements -> diverging -> exploring -> converging". But conversation is linear—ask question A, get result B. If you want to look back and do horizontal exploration, the dialog box becomes very clumsy. Conversation cannot accommodate the analyst's divergent exploration methods.

2. Mental Representation and Dimensionality Reduction Convergence

We believe that data analysis is essentially a modeling process. Humans have an internal "mental representation" of the world and business, which is difficult to verbalize. The analysis process uses means such as jotting down notes, dragging, SQL or code to gradually "reduce the dimensionality" of this complex mental representation, ultimately converging to a specific scenario, giving a chart, a conclusion or a decision reference.

What the user says is often the "simplified expression" after filtering countless ideas in their mind—it is not the truest thinking about data.

---

3. Technical Philosophy: AI's Boundaries and "Verifiability" Engineering

How to define AI's responsibilities in the system is crucial. AskTable proposes a clear set of engineering principles: AI applications are about engineering all "verifiable" things.

1. Distinguishing "Verifiable" from "Non-Verifiable"

• •Verifiable Logic (Verifiable Logic): Is the SQL syntax correct? Can the Python code run? Does the data calculation result conform to logic? These have standard answers and are areas where AI excels (Code Generation).
• •Non-Verifiable Logic (Unverifiable Logic): Intuition for business decisions, complex business background assumptions, preferences of a certain leader.

AskTable's strategy is: never let AI "guess" unverifiable things (like generating a fancy dashboard out of thin air), but use AI's extremely low marginal cost to solve all verifiable code generation and data processing tasks.

2. Stability Over Everything: Test-Driven Development (TDD)

In enterprise-level data scenarios, accuracy is the lifeline. AskTable emphasizes:

• •Importance of SQL Parser: Don't blindly go all-in on end-to-end model output. Must combine hand-written SQL parser (Parser) to define safety boundaries.
• •Evaluation Sets (Evals): Must establish automated evaluation sets for specific business scenarios. AI in data analysis is not "generative creation" but "functional execution," must test AI Agents like testing traditional software.

---

4. Exploring New Paradigms: From Agent to Non-Linear Canvas

To achieve "what you think is what you get," we conducted two core explorations in the product:

1. Agent Experimental Environment

Before exploring new products, we did an experiment, trying to place the Agent in a fusion environment integrating Python, SQL, DuckDB, VectorSearch, JavaScript and other tools. Give it a goal, let it try, feedback, iterate on its own, and generate in-depth reports containing thinking processes (Plan). Although this is not yet fully production-ready, it validated that AI, given tools and initial conditions, can give unexpectedly high-quality results.

2. Canvas

This is our recently internally tested core feature, designed to break the linear shackles.

2.1 Non-Linear Point-Edge Relationships: Nodes and Flows

Different from traditional BI's drag-and-drop or ChatBI's Q&A, AskTable's Canvas abstracts the analysis process into orchestratable nodes. The entire product is composed of "nodes," representing data flow through points and edges, rather than simple Q&A history.

• •Data Node: Carries raw data or query results, which may come from MySQL, Oracle, StarRocks, or even web pages.
• •Chart Node: Visual expression of data.
• •Computing Node (Python/SQL Node): This step is key. AskTable allows AI to write and execute Python code (like Pandas, Numpy processing) or SQL to handle complex logical operations.

2.2 Context Management

In Canvas, users can explicitly define AI's Context by drawing connections or selecting multiple nodes.

• •Scenario Example: User selects the data node of "last 20 orders," connects to a new Python node, and instructs "analyze these customers' profiles."
• •Technical Implementation: The system injects the preceding node's Schema, data summary, or even intermediate results as Context into the Prompt, achieving seamless transfer of data flow and logical flow.

This architecture turns the analysis process into a visual mind map. Users are no longer limited by programming languages or a single conversation window, but complete the mapping from "mental representation" to "formal expression" through node combinations.

---

5. Engineering Principles We Adhere To

In technical implementation, AskTable always follows these principles:

1. •Low Entropy Architecture: We continuously simplify the system architecture, evolving from the initial complex environment to lightweight deployment now, reducing implementation and operations entropy.
2. •Alignment with Standards: We don't arbitrarily create new terms and protocols unless there are extremely clear needs. We fully align with the benchmark protocols of the AI industry.
3. •Core Parser: Despite AI's power, we still value traditional engineering. We hand-wrote an SQL parser to parse user SQL and define boundaries, making AI's thinking safer, more reliable, and more controllable.

---

6. Conclusion: Invitation to Explore New Possibilities in Interaction

Writing an Agent is not difficult, but at this point in 2025, exploring new possibilities in human-AI interaction is the real challenge. We don't want to separate humans and AI, but to encapsulate trivial verifiable work within AI, freeing people for thinking exploration.

AI is responsible for tedious, verifiable logic, humans focus on abstract thinking and business exploration.

We are looking for partners who are sensitive to business and data to participate in our Canvas feature internal testing. If you have any ideas, or want to learn from our approach, welcome to exchange with us.

The AskTable team is committed to exploring the next frontier in human-AI collaboration.