Skip to main content
Version: PromptQL

Schema Selector

Overview​

The Schema Selector controls which DDN metadata (models, columns, and commands) are available to PromptQL. For large schemas, this optimizes performance, reduces costs, and improves accuracy by limiting the schema context sent to the LLM.

PromptQL provides three schema selection strategies:

Configuration Types​

All Schema (Default)​

Includes all available DDN metadata. This is the default behavior when schemaSelector is not provided.

All schema configuration
schemaSelector:
  type: all

For complete configuration details, see AllSchema reference.

Fixed Schema​

Explicitly specify which models and columns to include.

Fixed schema configuration
schemaSelector:
  type: fixed
  tables:
    - schemaName: app # Subgraph name
      tableName: users # Model name
      columns: [id, name, email, created_at]
    - schemaName: app
      tableName: orders
      columns: [id, user_id, total, status, created_at]
    - schemaName: analytics # Different subgraph
      tableName: user_metrics
      columns: [] # Empty array includes all columns

Note: schemaName refers to the subgraph name, tableName refers to the model name.

For complete configuration details, see FixedSchema reference.

LLM-Assisted Schema Selection​

Uses an LLM to intelligently select relevant schema elements for each user interaction. The selected schema is then provided to the PromptQL agent to fulfill the user's query.

LLM-assisted schema selector configuration
schemaSelector:
  type: llm_selector
  llm:
    provider: gemini
    model: gemini-2.5-flash
    apiKey:
      valueFromEnv: GEMINI_API_KEY
  maxItems: 25
  instructions: Focus on user-related models and analytics commands
  alwaysInclude:
    tables:
      app: [users, accounts] # Subgraph: [model names]
    functions: [calculate_user_stats] # Command names

Configuration options:

  • maxItems: Maximum number of models and commands to select (default: 25)
  • instructions: Custom guidance for the LLM selector
  • alwaysInclude: Schema elements to always include

For complete configuration details, see SchemaSelector reference.

This approach is ideal for large schemas where query patterns are unpredictable. It provides automatic optimization by dynamically selecting the most relevant schema elements for each user interaction.

LLM Context Requirements

For large schemas, the LLM used for schema selection must have a large context window to handle the full schema.

Next Steps​