Can AI Supercharge Low‑Code Dashboard Development? Lessons from Wuqi

AIGC is booming, and every field is experimenting. Does automatically generating a management dashboard with AI improve development efficiency? If AI can quickly develop a dashboard, can AI generation completely replace low‑code visual configuration? If AI cannot develop the entire dashboard, in which scenarios does AI work best?

Introduction

With the advent of ChatGPT, Wuqi has also integrated AI, hoping to further improve low‑code development efficiency.

Wuqi is a low‑code tool for management dashboards, originally designed to solve the development efficiency problem of internal enterprise dashboards . The capabilities of ChatGPT (referred to as AI below) suggest a new development method: can we complete dashboard development through human‑AI dialogue ? Considering AI's uncertainty, how should we integrate AI to achieve better efficiency?

Based on these questions, the Wuqi team explored for a period and made some staged progress; the following sections describe Wuqi’s incremental AI integration in the order of “Before AI” → “Partial AI Integration” → “Full‑Scene AI Integration” → “AI Customization”.

Before AI, Wuqi’s drag‑and‑drop was already refined

In low‑code platforms, the usual approach is to drag components first and then bind data. Wuqi introduces a new idea – "Intelligent UI Derivation" – where users only need to connect data or services to directly drag UI components.

AI arrives, Wuqi begins building an intelligent low‑code platform

Replacing a single drag‑and‑drop with AI makes no sense and can even be slower. Imagine that without AI, a list component can be dragged out with one click from data; with AI, you must type a long prompt (“Based on xxx data, generate a CRUD list component”), wait for the AI response, and may receive an unexpected component.

Therefore, not every scenario suits AI . We selected specific scenarios and gradually introduced AI to achieve the best effect.

Scenario 1: Component Customization

For a form component, after generating a form via Intelligent UI Derivation , further detailed adjustments are often repetitive and tedious. We tried using AI to replace these repetitive operations.

The video case demonstrates form label translation and dropdown option configuration . By using AI to replace simple repetitive manual operations, the low‑code platform’s usage threshold is lowered .

Scenario 2: Logic Customization

In Wuqi, some page logic requires a small amount of front‑end code (JavaScript), called LessCode . Within LessCode you can call APIs to "get page state data" or "invoke Wuqi utility methods". The code size is small and technical requirements are low, but the difficulty lies in learning Wuqi’s documentation – what if we give that documentation to AI?

The video case shows how to use AI in Wuqi to write logic code. After AI writes it, we can manually code‑review or skip the review and directly preview the result, reducing the user’s coding effort.

Scenario 3: Data Analysis

While AI assistance lowers the low‑code platform’s usage threshold in the previous two scenarios, it is still insufficient. Wuqi solves the end‑to‑end development from data to UI; can AI achieve full automation? We further attempt “zero‑code” development, allowing users to express requirements in a single sentence.

In the BI example, AI automatically writes SQL, selects appropriate chart components, and further adjusts component details; the whole process from "data integration" → "component generation" → "component adjustment" is completed independently by AI.

Full‑Scene AI Integration

Wuqi’s functions have now successfully integrated AI (partial AI). The key method is encapsulating a DSL layer :

Encapsulate DSL so AI understands Wuqi

First, according to atomic principles, wrap Wuqi’s capabilities into API interfaces , such as createData (create data) and createComp (drag component). This forms Wuqi’s DSL .

Then send these APIs and their parameter descriptions to AI, allowing AI to assemble API calls based on user instructions.

Finally, AI parses the user command into a sequence of API calls.

Wuqi as AI’s command‑line tool For AI‑generated DSL, Wuqi directly parses and executes it.

For example, on the left side of the image, the user inputs “generate a bar chart”, and AI outputs createData and createComp (with parameters). Wuqi executes them and a bar chart bound to user data appears on the canvas.

In some scenarios, predefined DSL alone cannot fulfill the user’s command, so AI also needs page state information. For instance, to “switch the chart’s axes”, AI first calls getCompInfo to obtain chart details, then uses setComp to modify it.

From a product perspective, Wuqi currently treats AI as an assistant tool . Users select a scenario and then use AI for rapid feature development. Could AI do more, such as page interaction design or visual optimization? For example, a user might ask AI to “generate a simple, beautiful login page”, and AI would handle data binding, form design, and login logic.

From a technical perspective, token limits are not yet a major issue, but as user commands become more complex, token limits could become critical . Additionally, AI may suffer from context interference, e.g., generating a bar chart when asked to modify a pie chart’s title.

Relying on a single AI instance cannot meet all needs; can multiple AI instances cooperate? We referenced AutoGPT and attempted to implement automatic task decomposition and automatic task dispatch/delivery , enabling multiple AI agents to collaboratively fulfill user commands.

Task Decomposition First, modularize the DSL into four modules corresponding to Wuqi’s capabilities: Navigation , Material , Canvas , Component Configuration . Any user instruction can be broken down into a combination of these modules. For example, “switch chart axes” requires fetching component info from Canvas and then modifying it via Component Configuration . This is akin to “Chain‑of‑Thought (CoT)”.

Task Dispatch and Delivery For the “switch chart axes” example, at least three AI agents are needed:

“Wuqi Grand Manager” handles task decomposition.

“Canvas Supervisor” handles Canvas functions.

“Component Config Supervisor” handles Component Configuration functions.

Communication between agents is achieved via a dedicated DSL, e.g., assign(fromAI, toAI, content) .

After solving task decomposition and delivery, recursive calls (similar to DFS) are needed. The DSL is hierarchical, forming a DSL tree. For example, the Material module can be further split into Data and Component Library, and Data can be split further, until leaf nodes are atomic APIs.

By merging any DSL node with its child nodes into a prompt node, we obtain a prompt tree that can generate multiple AI managers, forming an AI team that works for Wuqi users.

PS: The diagram explains how a DSL node merges into a prompt node: a prompt node consists of “the DSL node’s function”, “all child DSL nodes’ functions”, and optionally “the method to decompose child functions”.

Further video demonstrations will be released in the next Wuqi AI version.

For deeper understanding, here is a sample prompt for the Grand Manager and the AI’s JSON response:

# System context defining the Grand Manager’s basic rules and subordinate abilities
role: system
content: You are a website builder to help user develop website on a low‑code platform.
Constraints:
1. Write command arguments in JSON format.
2. Write thoughts in Chinese.
3. Use only the commands and agents below to solve the task.
4. When running runAgent, provide the information obtained from the current task in the task argument.
5. If you cannot solve the task, break it into smaller tasks.
6. Respond with JSON as described.
Commands:
1. taskComplete(): "run after the whole task is completed"
2. runAgent("agentId", "task"): "If the agent exists, run it to solve the task."
Resources:
1. You can use commands below to solve the task.
2. If the agent exists in Agents, you can use runAgent.
Agents:
material: ...
canvas: ...
component: ...
Response Format:
{ "thoughts": "...", "command": { "name": "...", "args": { "...": "..." } }, "next": true }

Beyond that, Wuqi can develop intelligent management‑dashboard applications

The AI capabilities should not be limited to the editor; users can build AI‑enabled applications on the Wuqi platform – "AI low‑code".

We added a "chatGPT" behavior atom (see Wuqi documentation) to allow users to configure various AI behaviors. Examples include AI‑filled forms and multi‑AI conversation demos.

Data Security Boundaries

Wuqi currently uses ChatGPT; here we explain data security boundaries.

As a management‑dashboard low‑code platform, the data it manipulates are internal enterprise data and must not be leaked. While the data itself is sensitive, the data structure and page structure are not sensitive .

For the "auto‑generate analysis chart" scenario, only the data schema is sent to AI, without actual data. For "AI auto‑extract article summary", the article is public, so sending it to AI is acceptable; however, for internal‑only articles, using ChatGPT is unsuitable.

Is AI the Most Efficient?

Based on the cases, AI is not always the most efficient. Efficiency depends on two factors:

1. Use Scenario

For a very specific small scenario, such as "generate a list from a fixed student data table", low‑code can generate the list with one click, making AI unnecessary. If the requirement changes to "calculate gender ratio from the student table", AI can understand natural language and generate SQL, outperforming low‑code.

Similarly, changing a form label from "name" to "名称" is faster manually, but if the operation repeats across many fields, AI’s batch processing shines.

Thus, suitable AI scenarios involve considerable thinking cost or large operation volume .

2. AI Maturity

Hallucinations are a fatal issue for large models; prompt engineering can mitigate but cannot eliminate occasional incorrect answers, which adds verification time.

Therefore, AI and humans will coexist for a long time. We can adopt a "human‑then‑AI" or "AI‑then‑human" workflow, using Wuqi’s intelligent UI for quick scaffolding and AI for detailed adjustments.

Can AI Replace Low‑Code?

After AI integration, many tasks can be completed via an "AI chat box", and its usage will increase. However, all AI‑generated outputs must be reviewed and adjusted by humans before production.

AI‑generated images, text, or videos often require post‑processing. Even if AI generates a full website code, developers still need to read, debug, and maintain it, which can be more labor‑intensive than using low‑code directly.

AI must be combined with tools; in the AI‑generated website domain, low‑code is the best tool. AI should feed low‑code, which then serves as an intermediate station for rapid verification and modification.

Everything Can Be AI, but AI Is Not the Only Answer

In summary, Wuqi’s AI integration abstracts a DSL layer, allowing AI to understand Wuqi’s capabilities and return DSL commands for execution. This abstraction enables any domain to integrate AI as long as its capabilities can be described atomically. The three steps are:

Encapsulate domain capabilities into DSL.

Refine DSL composition (Chain‑of‑Thought) into prompts.

Execute AI‑returned commands.

Wuqi’s AI journey will continue, aiming not only to boost efficiency but also to become a visual editing tool for AI‑generated management dashboards, fully merging low‑code and AI.