State

A State is the shared data dictionary that flows through your Pipeline. A State is currently defined as a TypedDict, so keys and value types are explicit. Each step reads from and writes to the State as it executes.

Default State: RAGState

By default, a Pipeline in our SDK is equipped with RAGState as it state, which is defined in gllm_pipeline.pipeline.states. The state keys are based on the keys that you might find in an Retrieval-Augmented Generation (RAG) pipeline. It is also equipped with a special state key for an EventEmitter for streaming purposes.

class RAGState(TypedDict):
    user_query: str
    queries: list[str]
    retrieval_params: dict[str, Any]
    chunks: list
    history: str
    context: str
    response: str
    references: str | list[str]
    event_emitter: EventEmitter

During Pipeline definition, you must ensure that your output state belongs to one of the pre-defined state keys.

Defining a custom State

The default RAGState may not be suitable for your purposes. For example, when defining a Subgraph, you may not need all of the keys. In contrast, there may be some additional state keys that you require. In these cases, you can define your own state structure.

To do so:

1

Create state class

Define a TypedDict . This can be in the same file as the Pipeline or in a different module.

my_state.py

from typing import TypedDict, Any
from gllm_core.event.event_emitter import EventEmitter

class MyCustomState(TypedDict):
    user_query: str
    chunks: list
    context: str
    response: str
    document_scores: list[float]  # not in RAGState
    debug_info: dict[str, Any]   # not in RAGState
    event_emitter: EventEmitter

2

Apply to pipeline

Pass the TypedDict into the state_type argument when creating the Pipeline.

from my_state import MyCustomState
from gllm_pipeline.pipeline import Pipeline

pipeline = Pipeline(
    steps=[
        retriever_step,
        response_synthesizer_step
    ],
    state_type=MyCustomState # 👈 custom field
)

Using a Pydantic BaseModel as a State

Coming Soon!