Implementation:Run llama Llama index LLMQuestionGenerator

Knowledge Sources	Run_llama_Llama_index
Domains	QuestionGeneration, LLM, SubQuestions
Last Updated	2026-02-11 19:00 GMT

Overview

LLMQuestionGenerator is a question generator that uses an LLM to decompose a complex query into a list of sub-questions, each targeting a specific tool, for multi-step retrieval and reasoning.

Description

LLMQuestionGenerator extends BaseQuestionGenerator and implements the sub-question generation pattern used in LlamaIndex's multi-step query decomposition. Given a complex query and a set of available tools (with their metadata), it prompts an LLM to break the query into focused sub-questions, each mapped to the most appropriate tool.

The generation workflow:

1. Available tools are formatted into a text description using build_tools_text.
2. The query string and tools text are passed to the LLM via the configured prompt template.
3. The LLM's response is parsed by the prompt's output_parser (defaults to SubQuestionOutputParser) to extract a structured list of SubQuestion objects.

Both synchronous (generate) and asynchronous (agenerate) methods are provided.

The class is typically instantiated via the from_defaults factory method, which handles default initialization of the LLM (from Settings.llm), prompt template (DEFAULT_SUB_QUESTION_PROMPT_TMPL with PromptType.SUB_QUESTION), and output parser (SubQuestionOutputParser). The constructor validates that the prompt has an output parser set.

Prompt management is implemented via _get_prompts and _update_prompts, exposing the prompt under the key "question_gen_prompt". When updating, the output parser is preserved or defaulted to SubQuestionOutputParser if missing.

Usage

Use LLMQuestionGenerator as part of a SubQuestionQueryEngine to automatically decompose complex queries that span multiple data sources or require multi-hop reasoning. Each generated sub-question targets a specific tool (e.g., a specific index or API), enabling parallel or sequential sub-query execution followed by answer synthesis.

Code Reference

Source Location

• Repository: Run_llama_Llama_index
• File: llama-index-core/llama_index/core/question_gen/llm_generators.py

Signature

class LLMQuestionGenerator(BaseQuestionGenerator):
    def __init__(
        self,
        llm: LLM,
        prompt: BasePromptTemplate,
    ) -> None:

    @classmethod
    def from_defaults(
        cls,
        llm: Optional[LLM] = None,
        prompt_template_str: Optional[str] = None,
        output_parser: Optional[BaseOutputParser] = None,
    ) -> "LLMQuestionGenerator":

Import

from llama_index.core.question_gen.llm_generators import LLMQuestionGenerator

I/O Contract

Inputs (from_defaults)

Name	Type	Required	Description
llm	LLM	No	LLM for generating sub-questions. Defaults to Settings.llm.
prompt_template_str	str	No	Custom prompt template string. Defaults to DEFAULT_SUB_QUESTION_PROMPT_TMPL.
output_parser	BaseOutputParser	No	Parser for extracting sub-questions from LLM output. Defaults to SubQuestionOutputParser.

Inputs (generate / agenerate)

Name	Type	Required	Description
tools	Sequence[ToolMetadata]	Yes	Metadata for available tools that sub-questions can target.
query	QueryBundle	Yes	The complex query to decompose into sub-questions.

Outputs

Name	Type	Description
sub_questions	List[SubQuestion]	List of generated sub-questions, each with a sub_question string and a tool_name identifying the target tool.

Usage Examples

from llama_index.core.question_gen.llm_generators import LLMQuestionGenerator
from llama_index.core.tools import ToolMetadata
from llama_index.core.schema import QueryBundle

# Create generator with defaults
question_gen = LLMQuestionGenerator.from_defaults()

# Define available tools
tools = [
    ToolMetadata(name="financial_docs", description="Financial reports and earnings data"),
    ToolMetadata(name="news_index", description="Recent news articles"),
]

# Generate sub-questions
query = QueryBundle(query_str="Compare the company's Q3 earnings with recent news coverage")
sub_questions = question_gen.generate(tools=tools, query=query)

for sq in sub_questions:
    print(f"Tool: {sq.tool_name}, Question: {sq.sub_question}")

Related Pages

• Environment:Run_llama_Llama_index_Python_LlamaIndex_Core
• Run_llama_Llama_index_Question_Gen_Types - Base classes and SubQuestion model
• Run_llama_Llama_index_PromptMixin - Prompt management mixin inherited via BaseQuestionGenerator