Configure Tool Instructions for Deep Dive Analysis

Open

The Gemini model mapping is not finalized in this version.

The Anthropic Bedrock model mapping is not finalized in this version.

• The Planner is responsible for determining how Deep Dive should answer a user query. It decides whether to use the Knowledge Base tool, SQL tool, metadata lookup, web search, code execution, or a combination of these tools.

• The Planner requires reasoning because it performs decision-making and orchestration. It must understand user intent, break down complex questions, select the appropriate tools, and generate a valid execution plan.

• Low reasoning is recommended because it provides a balance between planning quality, latency, and cost.

• A cost-optimized model is not recommended for the Planner because incorrect planning can affect the overall answer generation process.

Use Planner Instructions to:

• Define the preferred order of tool execution.

• Specify how ambiguous or incomplete questions should be handled.

• Control when the agent should use specific tools.

• Define planning rules for complex analytical queries.

Example

• Search the Knowledge Base before querying semantic models.

• Use Web Search only if the required information is not available in the configured data sources.

• Always validate results before generating the final response.

• The Summarizer is responsible for generating the final response after the required context, tool results, or intermediate outputs are available.

• The Summarizer is primarily responsible for synthesizing information, formatting the response, and ensuring clarity. It does not typically require multi-step reasoning once the required information is available.

• Reasoning should be disabled for this workflow to reduce latency and cost while maintaining response quality.

Use Summarizer Instructions to:

• Control the response format and level of detail.

• Specify the preferred writing style or tone.

• Define when to use tables, charts, or lists.

• Configure how recommendations, caveats, or assumptions are presented.

Example

• Present key findings as bullet points.

• Include a summary followed by supporting details.

• Highlight assumptions and limitations in the response.

Tool-specific instructions allow you to customize the behavior of individual tools available in the AI Space.

• SQL Generation is responsible for converting user questions into valid SQL queries based on the available semantic model, metadata, tables, columns, relationships, and business rules.

• The selected model should be able to:

  • Follow schema constraints.

  • Generate only one SQL statement, where required.

  • Respect tenant filters and security constraints.

  • Produce valid SQL for the target database dialect.

  • Avoid unsupported syntax.

  • Follow the required output format.

  • Avoid making assumptions when required metadata is missing.

• Reasoning is disabled by default because the SQL generation prompt should guide the model using clear rules, examples, schema context, and output constraints.

• GPT-4.1 may be used as a cost-optimized option when cost is a concern. However, SQL generation is sensitive to correctness, so customer-specific validation should be performed before using a cost-optimized model for production SQL workloads.

For example, you can configure instructions for the SQL Generator and Executor to influence how SQL queries are generated and executed.

• Knowledge Base Summarization is used to summarize retrieved knowledge base content, document chunks, document sections, or intermediate context before generating the final response.

• This is typically a high-volume internal workflow. The task focuses on concise summarization and preserving important information rather than performing deep reasoning.

• A smaller or cost-efficient model is suitable for this workflow, provided it preserves key information and generates consistent summaries.
  Example:

  • When asked about a KPI, return its threshold values alongside the data.

  • Recommend corrective actions based on approved best practices.

• Web Search is used to synthesize answers from retrieved web content or search results.

• This typically requires concise synthesis of the retrieved information rather than deep reasoning.

• The selected model should prioritize clear, source-grounded responses while maintaining low latency and cost efficiency.

The Instruction Tips pane provides guidance and examples for configuring each instruction type. Use these recommendations to write effective instructions that align the agent's planning, execution, and response generation with your organization's requirements.