Implementation:Run llama Llama index BaseSelector

Knowledge Sources	Run_llama_Llama_index
Domains	LLM Framework, Selection, Routing
Last Updated	2026-02-11 19:00 GMT

Overview

BaseSelector is the abstract base class for LLM-powered selector components that choose one or more options from a list of choices based on a query, used primarily in router query engines and retriever routing.

Description

This module defines the selection framework used for routing queries to appropriate sub-components. It contains several key classes and utilities:

Data models:

• SingleSelection -- A Pydantic model representing a single choice, containing an index (int) and a reason (str) explaining why the choice was selected.
• MultiSelection -- A Pydantic model containing a list of SingleSelection objects. Provides convenience properties:
  • ind -- Returns the single index (raises ValueError if more than one selection exists).
  • reason -- Returns the single reason (raises ValueError if more than one).
  • inds -- Returns a list of all selected indices.
  • reasons -- Returns a list of all selection reasons.
• SelectorResult -- An alias for MultiSelection, used for clarity in function signatures.

Helper functions:

• _wrap_choice(choice) -- Converts a MetadataType (either a string or ToolMetadata) into a ToolMetadata object.
• _wrap_query(query) -- Converts a QueryType (either a string or QueryBundle) into a QueryBundle object.

BaseSelector class:

• Inherits from PromptMixin and DispatcherSpanMixin.
• select(choices, query) -- Synchronous public method. Wraps choices and query, then delegates to abstract _select.
• aselect(choices, query) -- Async public method. Wraps choices and query, then delegates to abstract _aselect.
• _select and _aselect -- Abstract methods that subclasses must implement.

Usage

Subclass BaseSelector to create LLM-based selectors that route queries. This is used in RouterQueryEngine and RouterRetriever where the selector determines which sub-engine or sub-retriever should handle a given query based on descriptions of available choices.

Code Reference

Source Location

• Repository: Run_llama_Llama_index
• File: llama-index-core/llama_index/core/base/base_selector.py
• Lines: 1-103

Signature

class SingleSelection(BaseModel):
    index: int
    reason: str

class MultiSelection(BaseModel):
    selections: List[SingleSelection]
    @property
    def ind(self) -> int: ...
    @property
    def reason(self) -> str: ...
    @property
    def inds(self) -> List[int]: ...
    @property
    def reasons(self) -> List[str]: ...

SelectorResult = MultiSelection

class BaseSelector(PromptMixin, DispatcherSpanMixin):
    def select(
        self, choices: Sequence[MetadataType], query: QueryType
    ) -> SelectorResult: ...

    async def aselect(
        self, choices: Sequence[MetadataType], query: QueryType
    ) -> SelectorResult: ...

    @abstractmethod
    def _select(
        self, choices: Sequence[ToolMetadata], query: QueryBundle
    ) -> SelectorResult: ...

    @abstractmethod
    async def _aselect(
        self, choices: Sequence[ToolMetadata], query: QueryBundle
    ) -> SelectorResult: ...

Import

from llama_index.core.base.base_selector import (
    BaseSelector,
    SingleSelection,
    MultiSelection,
    SelectorResult,
)

I/O Contract

Inputs

Name	Type	Required	Description
choices	Sequence[MetadataType]	Yes	A sequence of choices, each being either a string description or a ToolMetadata object.
query	QueryType (str or QueryBundle)	Yes	The query used to determine which choice(s) to select.

Outputs

Name	Type	Description
return	SelectorResult (MultiSelection)	Contains a list of SingleSelection objects, each with an index and a reason string.

Usage Examples

Basic Usage

from llama_index.core.base.base_selector import BaseSelector, SelectorResult, SingleSelection
from llama_index.core.tools.types import ToolMetadata
from llama_index.core.schema import QueryBundle
from typing import Sequence

class MySelector(BaseSelector):
    def _get_prompts(self):
        return {}

    def _update_prompts(self, prompts):
        pass

    def _select(
        self, choices: Sequence[ToolMetadata], query: QueryBundle
    ) -> SelectorResult:
        # Simple keyword-based selection
        for i, choice in enumerate(choices):
            if query.query_str.lower() in choice.description.lower():
                return SelectorResult(
                    selections=[SingleSelection(index=i, reason="Keyword match")]
                )
        return SelectorResult(
            selections=[SingleSelection(index=0, reason="Default selection")]
        )

    async def _aselect(
        self, choices: Sequence[ToolMetadata], query: QueryBundle
    ) -> SelectorResult:
        return self._select(choices, query)

# Usage
selector = MySelector()
result = selector.select(
    choices=["Financial data analysis", "Scientific paper search"],
    query="What were Q3 earnings?"
)
print(result.ind, result.reason)

Related Pages

• Environment:Run_llama_Llama_index_Python_LlamaIndex_Core