Data Store Management

gllm-inference | Tutorial: Data Store Management | API Reference

Supported by: GoogleLMInvoker, OpenAILMInvoker

What is data store management?

Data store management is a feature that allows the language model to manage built-in data stores to be used as internal knowledge base. This allows the LM invoker to perform built-in RAG (Retrieval-Augmented Generation).

Data store management is only available for certain LM invokers. This feature can be accessed via the data_store attribute of the LM invoker. As an example, let's try to perform a simple built-in RAG using the GoogleLMInvoker!

Init an LM Invoker

First of all, let's create a GoogleLMInvoker that we will use to manage the data store:

from dotenv import load_dotenv
load_dotenv()

from gllm_inference.lm_invoker import GoogleLMInvoker

lm_invoker = GoogleLMInvoker("gemini-2.5-flash-lite")

Create a Data Store

Next, let's create a data store. The create() method will output an AttachmentStore object to be used in later operations.

store = await lm_invoker.data_store.create()

List the Data Stores

We can verify that the data store has been successfully created on the server side by using the list() method.

stores = await lm_invoker.data_store.list()

if not stores:
    print("No stores found.")

for store in stores:
    print(f" - {store}")

Add a File to the Data Store

Then, we can add a file to our newly created store using the add_file() method.

from gllm_inference.schema import Attachment

file = Attachment.from_path('path/to/file.pdf')
await lm_invoker.data_store.add_file(store, file)

Assign the Data Store to an LM invoker

Then, we can assign our store to the LM invoker to be used as an internal knowledge base.

lm_invoker.set_data_stores([store])

Alternatively, we can also directly assign the store to a new LM invoker:

lm_invoker = GoogleLMInvoker("gemini-2.5-flash-lite", data_stores=[stores])

During invocation, the LM invoker has the capability to retrieve knowledge from the stores that have been assigned to it, effectively enabling a built-in RAG.

output = await lm_invoker.invoke("<question about the file>")
print(f"output:\n{output}")

Delete the Data Store

Finally, if the store is no longer used, it can be deleted via the delete() method.

await lm_invoker.data_store.delete(store)