AutoRAG Custom Prompts Reference

Overview

AutoRAG allows you to customize the LLM prompts used during query processing. This enables you to:

• Tailor responses to your domain or use case
• Adjust the tone, style, or format of responses
• Provide domain-specific instructions to the LLM

Custom prompts are passed via the custom_prompts parameter in the query request. Only the prompts you specify are overridden; all others use their defaults.

Usage

{
  "query": "What are the main features of AI?",
  "query_type": 2,
  "custom_prompts": {
    "local_rag_response": "You are a technical expert. Provide a concise technical explanation focusing on implementation details.",
    "ds_generate_plan": "Break this query into 3-4 specific technical analysis steps."
  }
}

Prompts by query type

Each query type uses a specific set of prompts. The table below shows which prompts are available for which query type:

Prompt reference

Local and Unified prompts

• local_rag_response: Main prompt for generating responses in Local Search and Unified Search.

  • Used by: Local Search (query_type=2), Unified Search (query_type=3), when show_citations is true.
  • Template variables: {response_instructions}, {context_data}.

• local_rag_response_no_citations: Citation-free variant of local_rag_response.

  • Used by: Local Search and Unified Search when show_citations is false. The service selects this variant instead of appending a contradictory instruction to the citation-carrying template.
  • Template variables: same as local_rag_response.

Global prompts

• global_map_rag_points: Prompt for the map stage in Global Search. Processes individual community reports to extract key points with importance scores.

  • Used by: Global Search (query_type=1).
  • Template variable: {context_data}.

• global_reduce_rag_response: Prompt for the reduce stage in Global Search. Synthesizes ranked key points into a comprehensive final answer.

  • Used by: Global Search (query_type=1).
  • Template variables: {response_instructions}, {report_data}.

Deep Search prompts

These prompts are used when use_llm_planner=true with Local Search (query_type=2):

• ds_generate_plan: Generates a multi-step execution plan from the user query. Breaks down complex queries into 3-5 sequential steps.

  • Template variables: {global_context}, {original_query}.

• ds_step_query: Generates a specific, targeted query for each execution step. Fills in placeholders from previous step results.

  • Template variables: {step_description}, {expected_info}, {query_template}, {previous_results}.

• ds_completion_check: Evaluates whether the original query has been fully answered. Returns "SOLVED" or "CONTINUE".

  • Template variables: {original_query}, {execution_summary}.

• ds_final_synthesis: Synthesizes all step results into a comprehensive final answer.

  • Template variables: {original_query}, {response_instructions}, {all_step_results}.

• no_relevant_data_message: Message returned when no relevant data is found. Simple string (no template variables).

Custom Retriever Deep Search prompts

These prompts are used when use_llm_planner=true with Custom Retriever (query_type=4):

• ds_tool_selection: Selects the best tool for a query based on tool descriptions. Used in both pass 1 (custom_retriever tools) and pass 2 (local/global/unified tools).

  • Template variables: {tools_description}, {user_query}.

• ds_custom_generate_plan: Generates a multi-step execution plan tailored for Custom Retriever context.

  • Template variables: {global_context}, {original_query}.

• ds_custom_final_synthesis: Synthesizes all step results into a final answer for Custom Retriever queries.

  • Template variables: {original_query}, {response_instructions}, {all_step_results}.

• ds_cr_completion_check: Checks if the query has been fully answered after each step. Returns "SOLVED" or "CONTINUE".

  • Template variables: {original_query}, {execution_summary}.

• ds_no_tools_available: Message returned when no Custom Retriever tools are available. Simple string (no template variables).

Best practices

1. Start with defaults: Test with default prompts first to understand baseline behavior.
2. Incremental changes: Override one or two prompts at a time to see the impact.
3. Domain-specific instructions: Add domain-specific context to prompts (e.g., “You are a medical expert analyzing clinical data”).
4. Preserve template variables: When customizing prompts, keep the required {template_variables} so the system can inject context data.
5. Test across query types: Test custom prompts with various query types to ensure they work across different scenarios.

Validation

If you provide an invalid prompt key, the service returns an error listing the invalid keys and all valid keys:

{
  "error": "Invalid prompt keys found: ['invalid_key']. Valid keys are: ['ds_completion_check', 'ds_cr_completion_check', 'ds_custom_final_synthesis', 'ds_custom_generate_plan', 'ds_final_synthesis', 'ds_generate_plan', 'ds_no_tools_available', 'ds_step_query', 'ds_tool_selection', 'global_map_rag_points', 'global_reduce_rag_response', 'local_rag_response', 'local_rag_response_no_citations', 'no_relevant_data_message']"
}