Dialog2Graph

Dialog2Graph allows you to effortlessly create chatsky flows and scripts from dialogs using Large Language Models.

Contents

./dialog2graph - source code
./examples - usage scenarios
./experiments - test field for conducting experiments
./prompt_cache - utils for LLM output caching
./scripts - scripts for `poethepoet` automation 

Current Progress

Supported graph types:

• chain
• single cycle
• multi-cycle graph
• complex graph with cycles

Currently unsupported graph types:

• single node cycle

How to Start

Install poetry v. 1.8.4 (detailed installation guide)

pipx install poetry==1.8.4

Clone this repo and install project dependencies

git clone https://github.com/deeppavlov/dialog2graph.git
cd dialog2graph
poetry install

If you are planning to visualize your graphs consider installing PyGraphviz from here and also add it to the poetry environment.

poetry add pygraphviz

Ensure that dependencies were installed correctly by running any Python script

poetry run python <your_file_name>.py

Create .env file to store credentials

Note: never hardcode your personal tokens and other sensitive credentials. Use the .env file to store them.

How to Use

Generate synthetic graph on certain topic

Choose LLMs for generating and validating dialog graph and invoke graph generation

from dialog2graph.datasets.complex_dialogs.generation import LoopedGraphGenerator
from langchain_community.chat_models import ChatOpenAI


gen_model = ChatOpenAI(
    model='gpt-4o',
    api_key=os.getenv("OPENAI_API_KEY"),
    base_url=os.getenv("OPENAI_BASE_URL"),
)
val_model = ChatOpenAI(
    model='gpt-3.5-turbo',
    api_key=os.getenv("OPENAI_API_KEY"),
    base_url=os.getenv("OPENAI_BASE_URL"),
    temperature=0,
)

pipeline = LoopedGraphGenerator(
    generation_model=gen_model,
    validation_model=val_model,
)

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

Sample dialogs from existing dialog graph

Create graph instance and invoke sampler to get dialog list

from dialog2graph.pipelines.core.dialog_sampling import RecursiveDialogSampler
from dialog2graph.pipelines.core.graph import Graph

G = Graph(graph_dict={...})

sampler = RecursiveDialogSampler()
sampler.invoke(graph=G) #-> list of Dialog objects

How to Contribute

See contribution guideline CONTRIBUTING.md