Generate synthetic graph on certain topic

LoopedGraphGenerator is a class to create a validated graph from several LLM generated dialogs concerning a given topic. This class represents algorithm to generate dialog graph following several steps: graph creation from dialogs, graph validation evaluating theme consistency and cycle validity, dialog generation from graph.

from langchain_openai import ChatOpenAI

from dialog2graph.datasets.complex_dialogs.generation import LoopedGraphGenerator
from dialog2graph.pipelines.model_storage import ModelStorage

First, ModelStorage instance is created. Then, using add method, choosen LLMs may be added for dialog generation, dialog validation, theme validation and cycle end search, important generation steps. You can choose between available models. Preferably, the stronger model is used for generation problem and the simpler models are used for other tasks.

model_storage = ModelStorage()
model_storage.add(
    "gen_model", # model to use for generation
    config={"model_name": "o1-mini"},
    model_type=ChatOpenAI,
)
model_storage.add(
    "help_model", # model to use for other tasks
    config={"model_name": "gpt-3.5-turbo"},
    model_type=ChatOpenAI,
)

Then, LoopedGraphGenerator is created, passing previously defined ModelStorage and specifying model names for each task. invoke method is used to get a dialog graph on restaurant reservation topic. The method returns a dictionary ("graph": {{"edges": [...], "nodes": [...]}, "dialogues": [...]}) that can be passed to Graph (more info you can find in this userguide).

pipeline = LoopedGraphGenerator(
    model_storage=model_storage,
    generation_llm='gen_model',
    validation_llm='help_model',
    cycle_ends_llm='help_model',
    theme_validation_llm='help_model'
)

generated_graph = pipeline.invoke(topic="restaurant reservation")