Implementation:BerriAI Litellm Vector Store Registry

Property	Value
sources	litellm/vector_stores/vector_store_registry.py
domains	Vector Stores, Registry, Configuration, Database, Credentials
last_updated	2026-02-15 16:00 GMT

Overview

The Vector Store Registry module provides two registry classes that manage vector store configurations and indexes in-memory with database fallback, handling credential resolution, YAML config loading, tool parameter extraction, and cache synchronization across multiple proxy instances.

Description

This module contains two registry classes. VectorStoreIndexRegistry manages LiteLLM_ManagedVectorStoreIndex objects with methods for CRUD operations (get_vector_store_indexes, get_vector_store_index_by_name, upsert_vector_store_index, delete_vector_store_index, is_vector_store_index) and a static database helper (_get_vector_store_indexes_from_db) for loading indexes from the Prisma database.

VectorStoreRegistry is the primary registry class that manages LiteLLM_ManagedVectorStore objects. It maintains an in-memory list of vector stores with a vector_store_ids_to_vector_store_map for fast lookups. Key capabilities include:

• Configuration loading via load_vector_stores_from_config() which parses YAML config entries
• Tool parameter extraction via get_and_pop_recognised_vector_store_tools() which identifies vector store tools from request tool lists, extracts their parameters, removes them from the tool list, and returns a parameter-by-ID mapping
• Pop-and-run semantics via pop_vector_stores_to_run() and pop_vector_stores_to_run_with_db_fallback() which extract vector store IDs, merge tool parameters, and return copies with merged litellm_params
• Database synchronization via get_litellm_managed_vector_store_from_registry_or_db() and pop_vector_stores_to_run_with_db_fallback() which verify in-memory entries against the database and auto-remove stale entries
• Credential resolution via get_credentials_for_vector_store() which looks up litellm_credential_name and resolves it through CredentialAccessor

Usage

Import this module when you need to manage vector store configurations in the LiteLLM proxy, load vector stores from YAML config, resolve vector store credentials for API calls, or extract vector store tool parameters from request payloads. The VectorStoreRegistry is typically instantiated as litellm.vector_store_registry and used by the vector store search and file management modules.

Code Reference

Source Location

Property	Value
Repository	github.com/BerriAI/litellm
File	litellm/vector_stores/vector_store_registry.py
Lines	559
Module	litellm.vector_stores.vector_store_registry

Signature

class VectorStoreIndexRegistry:
    def __init__(self, vector_store_indexes: List[LiteLLM_ManagedVectorStoreIndex] = [])
    def get_vector_store_indexes(self) -> List[LiteLLM_ManagedVectorStoreIndex]
    def get_vector_store_index_by_name(self, name: str) -> Optional[LiteLLM_ManagedVectorStoreIndex]
    def upsert_vector_store_index(self, index: LiteLLM_ManagedVectorStoreIndex) -> None
    def delete_vector_store_index(self, index: str) -> None
    def is_vector_store_index(self, name: str) -> bool
    @staticmethod
    async def _get_vector_store_indexes_from_db(prisma_client) -> List[LiteLLM_ManagedVectorStoreIndex]

class VectorStoreRegistry:
    def __init__(self, vector_stores: List[LiteLLM_ManagedVectorStore] = [])
    def pop_vector_stores_to_run(self, non_default_params: Dict, tools: Optional[List[Dict]] = None) -> List[LiteLLM_ManagedVectorStore]
    async def pop_vector_stores_to_run_with_db_fallback(self, non_default_params: Dict, tools=None, prisma_client=None) -> List[LiteLLM_ManagedVectorStore]
    def get_and_pop_recognised_vector_store_tools(self, tools=None, vector_store_ids=None) -> Dict[str, VectorStoreToolParams]
    def load_vector_stores_from_config(self, vector_stores_config: List[Dict]) -> None
    def get_credentials_for_vector_store(self, vector_store_id: str) -> Dict[str, Any]
    def list_all_vector_stores(self) -> LiteLLM_ManagedVectorStoreListResponse
    def add_vector_store_to_registry(self, vector_store: LiteLLM_ManagedVectorStore) -> None
    def delete_vector_store_from_registry(self, vector_store_id: str) -> None
    def update_vector_store_in_registry(self, vector_store_id: str, updated_data: LiteLLM_ManagedVectorStore) -> None

Import

from litellm.vector_stores.vector_store_registry import (
    VectorStoreRegistry,
    VectorStoreIndexRegistry,
)

I/O Contract

Inputs

Method	Key Parameters	Description
load_vector_stores_from_config	vector_stores_config: List[Dict]	Parses YAML config entries with vector_store_name and litellm_params
pop_vector_stores_to_run	non_default_params: Dict, tools: Optional[List[Dict]]	Extracts and pops vector store IDs from params and tool definitions
get_credentials_for_vector_store	vector_store_id: str	Looks up credentials by vector store ID
get_and_pop_recognised_vector_store_tools	tools: Optional[List[Dict]]	Identifies, extracts, and removes vector store tools from tool list

Outputs

Method	Return Type	Description
pop_vector_stores_to_run	List[LiteLLM_ManagedVectorStore]	Vector stores with merged tool parameters in litellm_params
get_credentials_for_vector_store	Dict[str, Any]	Unpacked credential values for the vector store
list_all_vector_stores	LiteLLM_ManagedVectorStoreListResponse	Paginated response with all registered vector stores
get_and_pop_recognised_vector_store_tools	Dict[str, VectorStoreToolParams]	Mapping of vector store ID to extracted tool parameters

Usage Examples

from litellm.vector_stores.vector_store_registry import VectorStoreRegistry

# Initialize and load from config
registry = VectorStoreRegistry()
registry.load_vector_stores_from_config([
    {
        "vector_store_name": "knowledge_base",
        "litellm_params": {
            "vector_store_id": "vs_abc123",
            "custom_llm_provider": "openai",
            "api_key": "sk-...",
        },
    }
])

# Get credentials for a vector store
credentials = registry.get_credentials_for_vector_store("vs_abc123")

# Pop vector stores to run from a request
vector_stores = registry.pop_vector_stores_to_run(
    non_default_params={"vector_store_ids": ["vs_abc123"]},
    tools=[{"vector_store_ids": ["vs_def456"], "max_num_results": 5}],
)
for vs in vector_stores:
    print(f"Running vector store: {vs['vector_store_id']}")

Related Pages

• BerriAI_Litellm_Vector_Stores_API -- Vector store create and search operations that use this registry for credential lookup
• BerriAI_Litellm_Vector_Store_Files_API -- File management that uses registry credentials via _prepare_registry_credentials()