Supported Datastores

Overview

The GLLM DataStore library implements the Single Abstraction Pattern, meaning you interact with a unified API regardless of the underlying storage engine. Whether you use Chroma for prototyping or Elasticsearch for production, the code to store and retrieve data remains generic:

# The only difference is the initialization
store = ChromaDataStore(...)
# OR
store = ElasticsearchDataStore(...)

# The rest of your code stays the same
await store.vector.retrieve(query="...")

Feature Matrix

Datastore	Fulltext	Vector	Hybrid	Notes
Chroma	✅	✅	❌	Great for local dev & prototyping.
Elasticsearch	✅	✅	✅	Production-grade search engine.
OpenSearch	✅	✅	✅	Open-source alternative to ES.
Milvus	✅	✅	✅	High-performance vector database.
PostgreSQL	✅	✅	❌	Uses pgvector for vector support.
Redis	✅	✅	❌	In-memory speed with persistence.
SQL	✅	❌	❌	Generic SQL support (Postgres, MySQL, SQLite).
In-Memory	✅	✅	❌	For testing and ephemeral data.

Initialization Examples

To inspect the specific prerequisites for each datastore (e.g. pip install ...), please refer to the Optional Dependencies section.

Chroma

from gllm_datastore.data_store import ChromaDataStore
from gllm_datastore.data_store.chroma.data_store import ChromaClientType
from gllm_inference.em_invoker import OpenAIEMInvoker

# 1. Initialize Invoker (for vector search)
em_invoker = OpenAIEMInvoker(model_name="text-embedding-3-small")

# 2. Initialize Store
store = (
    ChromaDataStore(
        collection_name="my_collection",
        client_type=ChromaClientType.PERSISTENT,
        persist_directory="./chroma-data",
    )
    .with_fulltext()
    .with_vector(em_invoker=em_invoker)
)

Elasticsearch

OpenSearch

Milvus

PostgreSQL

Redis

SQL

In-Memory