Modeling (MODEL_GENERATION)

This section describes the end-to-end flow from chat2 to ai-link when the user asks for “modeling”-type requirements.

Typical utterances:

• “Create a customer model with name, age, and email.”
• “Design an order table with amount, customer, and order time.”

0. Prompt

You are a metadata modeling assistant. You must output only a JSON array.

Return structure:
[{
  "entity": "ai_orders",
  "description": "Orders",
  "fields": [
    {
      "name": "customer_name",
      "type": "string",
      "label": "Customer Name"
    }
  ]
}]

Constraints:
1. Output JSON only, no explanation or markdown
2. type must be one of: "string","number","date","enum"
3. For enum fields, if there are options, use the choices array
4. name uses snake_case, for example "customer_name"
5. label uses short Chinese or English titles
6. entity uses table-style names, such as "ai_customers","ai_orders"

1. Intent detection

In workflow ai-chat.json:

• Node start_chat receives the conversation message array messages.
• Node code_parse:
  • extracts the latest user message as query
  • builds a history string
  • analyzes context keywords and derives a high-level intent such as MODEL_GENERATION

Key logic (paraphrased):

• If keywords like “建模, 模型, 表结构, 字段, schema, 实体” appear, intent is judged as MODEL_GENERATION.

2. Intent routing

In the condition_intent node, routing is based on code_parse.intent:

• Branch port keys include:
  • MODEL_GENERATION
  • DATA_QUERY
  • DATA_MUTATION
  • DATA_ANALYSIS
  • PERMISSION_AUDIT
  • EXAMPLE_UI

When intent is MODEL_GENERATION, the workflow takes the “modeling” branch.

3. LLM generates model draft

In the modeling branch, an LLM node (for example llm_model) is called:

• provider / modelName / apiHost / apiKey are configured on the node
• systemPrompt forces the LLM to output only a JSON array
• Sample output:

[
  {
    "entity": "ai_customers",
    "description": "Customers",
    "fields": [
      { "name": "name", "type": "string", "label": "Name" },
      { "name": "age", "type": "number", "label": "Age" },
      { "name": "email", "type": "string", "label": "Email" }
    ]
  }
]

4. Frontend renders the model form

In subsequent code nodes, the model JSON is converted into a UI Schema that the frontend can render, and returned to chat2 as a tool card.

In ai-chat.json, there is a script (paraphrased):

• When intent === 'MODEL_GENERATION' and modelSchema exists:
  • build a form description object containing entity, fields, initialValues, and save actions
  • return it in a tool call:
    • toolName: "modelForm"
    • result: the form definition

The chat2 frontend uses toolName: "modelForm" to render a modeling form component, allowing the user to:

• adjust fields (name, type, label, etc.)
• click save/submit to send the final modeling request to the backend (typically Looker metadata APIs)

5. Corresponding APIs and persistence

When the user confirms the model in the form:

• The frontend calls Looker metadata APIs:
  • /metadata/entities
  • /metadata/fields
• It persists the LLM-generated structure plus user edits as real entities/fields, completing the modeling process.