"""
Tool for performing subgraph extraction.
"""
from typing import Type, Annotated
import logging
import pickle
import numpy as np
import pandas as pd
import hydra
import networkx as nx
from pydantic import BaseModel, Field
from langchain.chains.retrieval import create_retrieval_chain
from langchain.chains.combine_documents import create_stuff_documents_chain
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.vectorstores import InMemoryVectorStore
from langchain_core.tools import BaseTool
from langchain_core.messages import ToolMessage
from langchain_core.tools.base import InjectedToolCallId
from langchain_community.document_loaders import PyPDFLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langgraph.types import Command
from langgraph.prebuilt import InjectedState
import torch
from torch_geometric.data import Data
from ..utils.extractions.pcst import PCSTPruning
from ..utils.embeddings.ollama import EmbeddingWithOllama
from .load_arguments import ArgumentData
# Initialize logger
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
class SubgraphExtractionInput(BaseModel):
"""
SubgraphExtractionInput is a Pydantic model representing an input for extracting a subgraph.
Args:
prompt: Prompt to interact with the backend.
tool_call_id: Tool call ID.
state: Injected state.
arg_data: Argument for analytical process over graph data.
"""
tool_call_id: Annotated[str, InjectedToolCallId] = Field(
description="Tool call ID."
)
state: Annotated[dict, InjectedState] = Field(description="Injected state.")
prompt: str = Field(description="Prompt to interact with the backend.")
arg_data: ArgumentData = Field(
description="Experiment over graph data.", default=None
)
class SubgraphExtractionTool(BaseTool):
"""
This tool performs subgraph extraction based on user's prompt by taking into account
the top-k nodes and edges.
"""
name: str = "subgraph_extraction"
description: str = "A tool for subgraph extraction based on user's prompt."
args_schema: Type[BaseModel] = SubgraphExtractionInput
def perform_endotype_filtering(
self,
prompt: str,
state: Annotated[dict, InjectedState],
cfg: hydra.core.config_store.ConfigStore,
) -> str:
"""
Perform endotype filtering based on the uploaded files and prepare the prompt.
Args:
prompt: The prompt to interact with the backend.
state: Injected state for the tool.
cfg: Hydra configuration object.
"""
# Loop through the uploaded files
all_genes = []
for uploaded_file in state["uploaded_files"]:
if uploaded_file["file_type"] == "endotype":
# Load the PDF file
docs = PyPDFLoader(file_path=uploaded_file["file_path"]).load()
# Split the text into chunks
splits = RecursiveCharacterTextSplitter(
chunk_size=cfg.splitter_chunk_size,
chunk_overlap=cfg.splitter_chunk_overlap,
).split_documents(docs)
# Create a chat prompt template
prompt_template = ChatPromptTemplate.from_messages(
[
("system", cfg.prompt_endotype_filtering),
("human", "{input}"),
]
)
qa_chain = create_stuff_documents_chain(
state["llm_model"], prompt_template
)
rag_chain = create_retrieval_chain(
InMemoryVectorStore.from_documents(
documents=splits, embedding=state["embedding_model"]
).as_retriever(
search_type=cfg.retriever_search_type,
search_kwargs={
"k": cfg.retriever_k,
"fetch_k": cfg.retriever_fetch_k,
"lambda_mult": cfg.retriever_lambda_mult,
},
),
qa_chain,
)
results = rag_chain.invoke({"input": prompt})
all_genes.append(results["answer"])
# Prepare the prompt
if len(all_genes) > 0:
prompt = " ".join(
[prompt, cfg.prompt_endotype_addition, ", ".join(all_genes)]
)
return prompt
def prepare_final_subgraph(self,
subgraph: dict,
pyg_graph: Data,
textualized_graph: pd.DataFrame) -> dict:
"""
Prepare the subgraph based on the extracted subgraph.
Args:
subgraph: The extracted subgraph.
pyg_graph: The PyTorch Geometric graph.
textualized_graph: The textualized graph.
Returns:
A dictionary containing the PyG graph, NetworkX graph, and textualized graph.
"""
# print(subgraph)
# Prepare the PyTorch Geometric graph
mapping = {n: i for i, n in enumerate(subgraph["nodes"].tolist())}
pyg_graph = Data(
# Node features
x=pyg_graph.x[subgraph["nodes"]],
node_id=np.array(pyg_graph.node_id)[subgraph["nodes"]].tolist(),
node_name=np.array(pyg_graph.node_id)[subgraph["nodes"]].tolist(),
enriched_node=np.array(pyg_graph.enriched_node)[subgraph["nodes"]].tolist(),
num_nodes=len(subgraph["nodes"]),
# Edge features
edge_index=torch.LongTensor(
[
[
mapping[i]
for i in pyg_graph.edge_index[:, subgraph["edges"]][0].tolist()
],
[
mapping[i]
for i in pyg_graph.edge_index[:, subgraph["edges"]][1].tolist()
],
]
),
edge_attr=pyg_graph.edge_attr[subgraph["edges"]],
edge_type=np.array(pyg_graph.edge_type)[subgraph["edges"]].tolist(),
relation=np.array(pyg_graph.edge_type)[subgraph["edges"]].tolist(),
label=np.array(pyg_graph.edge_type)[subgraph["edges"]].tolist(),
enriched_edge=np.array(pyg_graph.enriched_edge)[subgraph["edges"]].tolist(),
)
# Networkx DiGraph construction to be visualized in the frontend
nx_graph = nx.DiGraph()
for n in pyg_graph.node_name:
nx_graph.add_node(n)
for i, e in enumerate(
[
[pyg_graph.node_name[i], pyg_graph.node_name[j]]
for (i, j) in pyg_graph.edge_index.transpose(1, 0)
]
):
nx_graph.add_edge(
e[0],
e[1],
relation=pyg_graph.edge_type[i],
label=pyg_graph.edge_type[i],
)
# Prepare the textualized subgraph
textualized_graph = (
textualized_graph["nodes"].iloc[subgraph["nodes"]].to_csv(index=False)
+ "\n"
+ textualized_graph["edges"].iloc[subgraph["edges"]].to_csv(index=False)
)
return {
"graph_pyg": pyg_graph,
"graph_nx": nx_graph,
"graph_text": textualized_graph,
}
def _run(
self,
tool_call_id: Annotated[str, InjectedToolCallId],
state: Annotated[dict, InjectedState],
prompt: str,
arg_data: ArgumentData = None,
) -> Command:
"""
Run the subgraph extraction tool.
Args:
tool_call_id: The tool call ID for the tool.
state: Injected state for the tool.
prompt: The prompt to interact with the backend.
arg_data (ArgumentData): The argument data.
Returns:
Command: The command to be executed.
"""
logger.log(logging.INFO, "Invoking subgraph_extraction tool")
# Load hydra configuration
with hydra.initialize(version_base=None, config_path="../configs"):
cfg = hydra.compose(
config_name="config", overrides=["tools/subgraph_extraction=default"]
)
cfg = cfg.tools.subgraph_extraction
# Retrieve source graph from the state
initial_graph = {}
initial_graph["source"] = state["dic_source_graph"][-1] # The last source graph as of now
# logger.log(logging.INFO, "Source graph: %s", source_graph)
# Load the knowledge graph
with open(initial_graph["source"]["kg_pyg_path"], "rb") as f:
initial_graph["pyg"] = pickle.load(f)
with open(initial_graph["source"]["kg_text_path"], "rb") as f:
initial_graph["text"] = pickle.load(f)
# Prepare prompt construction along with a list of endotypes
if len(state["uploaded_files"]) != 0 and "endotype" in [
f["file_type"] for f in state["uploaded_files"]
]:
prompt = self.perform_endotype_filtering(prompt, state, cfg)
# Prepare embedding model and embed the user prompt as query
query_emb = torch.tensor(
EmbeddingWithOllama(model_name=cfg.ollama_embeddings[0]).embed_query(prompt)
).float()
# Prepare the PCSTPruning object and extract the subgraph
# Parameters were set in the configuration file obtained from Hydra
subgraph = PCSTPruning(
state["topk_nodes"],
state["topk_edges"],
cfg.cost_e,
cfg.c_const,
cfg.root,
cfg.num_clusters,
cfg.pruning,
cfg.verbosity_level,
).extract_subgraph(initial_graph["pyg"], query_emb)
# Prepare subgraph as a NetworkX graph and textualized graph
final_subgraph = self.prepare_final_subgraph(
subgraph, initial_graph["pyg"], initial_graph["text"]
)
# Prepare the dictionary of extracted graph
dic_extracted_graph = {
"name": arg_data.extraction_name,
"tool_call_id": tool_call_id,
"graph_source": initial_graph["source"]["name"],
"topk_nodes": state["topk_nodes"],
"topk_edges": state["topk_edges"],
"graph_dict": {
"nodes": list(final_subgraph["graph_nx"].nodes(data=True)),
"edges": list(final_subgraph["graph_nx"].edges(data=True)),
},
"graph_text": final_subgraph["graph_text"],
"graph_summary": None,
}
# Prepare the dictionary of updated state
dic_updated_state_for_model = {}
for key, value in {
"dic_extracted_graph": [dic_extracted_graph],
}.items():
if value:
dic_updated_state_for_model[key] = value
# Return the updated state of the tool
return Command(
update=dic_updated_state_for_model | {
# update the message history
"messages": [
ToolMessage(
content=f"Subgraph Extraction Result of {arg_data.extraction_name}",
tool_call_id=tool_call_id,
)
],
}
)