kig_test / ki_gen /planner.py
adrienbrdne's picture
Update ki_gen/planner.py
77b16df verified
import os
import re
from typing import Annotated
from typing_extensions import TypedDict
# Remove ChatGroq import
# from langchain_groq import ChatGroq
# Add ChatGoogleGenerativeAI import
from langchain_google_genai import ChatGoogleGenerativeAI
import os # Add os import
from langchain_openai import ChatOpenAI
from langchain_core.messages import SystemMessage, HumanMessage
from langchain_community.graphs import Neo4jGraph
from langgraph.graph import StateGraph
from langgraph.graph import add_messages
from ki_gen.prompts import PLAN_GEN_PROMPT, PLAN_MODIFICATION_PROMPT
from ki_gen.data_retriever import build_data_retriever_graph
from ki_gen.data_processor import build_data_processor_graph
# Import get_model which now handles Gemini
from ki_gen.utils import ConfigSchema, State, HumanValidationState, DocProcessorState, DocRetrieverState, get_model
from langgraph.checkpoint.sqlite import SqliteSaver
##########################################################################
###### NODES DEFINITION ######
##########################################################################
def validate_node(state: State):
"""
This node inserts the plan validation prompt.
"""
prompt = """System : You only need to focus on Key Issues, no need to focus on solutions or stakeholders yet and your plan should be concise.
If needed, give me an updated plan to follow this instruction. If your plan already follows the instruction just say "My plan is correct"."""
output = HumanMessage(content=prompt)
return {"messages" : [output]}
# Remove Groq-specific error handler
# def error_chatbot_groq(error, model_name, query): ...
# Wrappers to call LLMs on the state messsages field
# Simplify: Use get_model directly or a single chatbot function
def chatbot_node(state: State, config: ConfigSchema):
"""Generic chatbot node using the main_llm from config."""
model_name = config["configurable"].get("main_llm") or "gemini-2.0-flash"
llm = get_model(model_name)
try:
# Check if messages exist and are not empty
if "messages" in state and state["messages"]:
response = llm.invoke(state["messages"])
return {"messages": [response]}
else:
print("Warning: No messages found in state for chatbot_node.")
# Return state unchanged or an empty message list?
return {} # Or {"messages": []}
except Exception as e:
print(f"Error invoking model {model_name}: {e}")
# Handle error, maybe return an error message or empty dict
return {"messages": [SystemMessage(content=f"Error during generation: {e}")]}
# Remove old chatbot functions (chatbot_llama, chatbot_mixtral, chatbot_openai)
# Replace the chatbots dictionary with direct calls to the generic function or specific models via get_model
# This simplifies planner.py, relying on utils.py and config for model selection.
def parse_plan(state: State):
"""
This node parses the generated plan and writes in the 'store_plan' field of the state
"""
# Find the AI message likely containing the plan (often the second to last if validate_node was used)
plan_message_content = ""
if "messages" in state and len(state["messages"]) >= 1:
# Search backwards for the plan, as its position might vary
for msg in reversed(state["messages"]):
if hasattr(msg, 'content') and "Plan:" in msg.content and "<END_OF_PLAN>" in msg.content:
plan_message_content = msg.content
break # Found the plan
if not plan_message_content:
print("Error: Could not find plan message in state.")
# Handle error: maybe return current state or raise an exception
return state # Return unchanged state if plan not found
store_plan = []
try:
# Improved parsing: handle potential variations in formatting
plan_section = plan_message_content.split("Plan:")[1].split("<END_OF_PLAN>")[0]
# Split by numbered steps, removing empty entries
store_plan = [step.strip() for step in re.split(r"\n\s*\d+\.\s*", plan_section) if step.strip()]
except Exception as e:
print(f"Error while parsing plan: {e}")
# Handle parsing error, potentially keep store_plan empty or log the error
store_plan = [] # Reset plan on error
return {"store_plan" : store_plan}
# Update get_detailed_query to use get_model and default model
def get_detailed_query(context : list, model : str = "gemini-2.0-flash"):
"""
Simple helper function for the detail_step node
"""
llm = get_model(model) # Use get_model
try:
return llm.invoke(context)
except Exception as e:
print(f"Error in get_detailed_query with model {model}: {e}")
# Return a default message or raise error
return SystemMessage(content=f"Error generating detailed query: {e}")
def detail_step(state: State, config: ConfigSchema):
"""
This node updates the value of the 'current_plan_step' field and defines the query to be used for the data_retriever.
"""
print("Entering detail_step") # Debug print
print(f"Current state keys: {state.keys()}") # Debug print
# Initialize current_plan_step if not present
current_plan_step = state.get("current_plan_step", -1) + 1
# Ensure store_plan exists and has enough steps
store_plan = state.get("store_plan", [])
if not store_plan or current_plan_step >= len(store_plan):
print(f"Warning: Plan step {current_plan_step} out of bounds or plan is empty.")
# Decide how to handle: end graph, return error state?
# For now, let's prevent index error and maybe signal an issue
# Returning an empty query might halt progress or cause issues downstream
return {"current_plan_step": current_plan_step, 'query' : "Error: Plan step unavailable.", "valid_docs" : []}
plan_step_description = store_plan[current_plan_step]
if config["configurable"].get("use_detailed_query"):
prompt = HumanMessage(f"""Specify what additional information you need to proceed with the next step of your plan :
Step {current_plan_step + 1} : {plan_step_description}""")
# Ensure messages exist before appending
current_messages = state.get("messages", [])
query_message = get_detailed_query(context = current_messages + [prompt], model=config["configurable"].get("main_llm", "gemini-2.0-flash"))
query_content = query_message.content if hasattr(query_message, 'content') else "Error: Could not get detailed query content."
return {"messages" : [prompt, query_message], "current_plan_step": current_plan_step, 'query' : query_content, "valid_docs": state.get("valid_docs", [])} # Ensure valid_docs is preserved
# If not using detailed query, use the plan step description directly
return {"current_plan_step": current_plan_step, 'query' : plan_step_description, "valid_docs" : state.get("valid_docs", [])} # Ensure valid_docs is preserved
def concatenate_data(state: State):
"""
This node concatenates all the data that was processed by the data_processor and inserts it in the state's messages
"""
# Ensure valid_docs exists and current_plan_step is valid
valid_docs_content = state.get("valid_docs", "No processed documents available.")
current_plan_step = state.get("current_plan_step", -1)
store_plan = state.get("store_plan", [])
if current_plan_step < 0 or current_plan_step >= len(store_plan):
print(f"Warning: Invalid current_plan_step ({current_plan_step}) in concatenate_data.")
# Handle error - maybe return an error message
step_description = "Error: Current plan step invalid."
else:
step_description = store_plan[current_plan_step]
prompt = f"""#########TECHNICAL INFORMATION ############
{str(valid_docs_content)}
########END OF TECHNICAL INFORMATION#######
Using the information provided above, proceed with step {current_plan_step + 1} of your plan :
{step_description}
"""
return {"messages": [HumanMessage(content=prompt)]}
def human_validation(state: HumanValidationState) -> HumanValidationState:
"""
Dummy node to interrupt before processing, can be used for manual validation later.
"""
# Defaulting to no processing steps needed unless specified elsewhere
return {'process_steps' : state.get('process_steps', [])}
def generate_ki(state: State):
"""
This node inserts the prompt to begin Key Issues generation
"""
print(f"THIS IS THE STATE FOR CURRENT PLAN STEP IN GENERATE_KI : {state.get('current_plan_step')}")
current_plan_step = state.get("current_plan_step", -1)
store_plan = state.get("store_plan", [])
# Check if the next step exists in the plan
next_step_index = current_plan_step + 1
if next_step_index < 0 or next_step_index >= len(store_plan):
print(f"Warning: Invalid next plan step ({next_step_index}) for KI generation.")
step_description = "Error: Plan step for KI generation unavailable."
else:
step_description = store_plan[next_step_index]
prompt = f"""Using the information provided above, proceed with step {next_step_index + 1} of your plan to provide the user with NEW and INNOVATIVE Key Issues :
{step_description}"""
return {"messages" : [HumanMessage(content=prompt)]}
def detail_ki(state: State):
"""
This node inserts the last prompt to detail the generated Key Issues
"""
current_plan_step = state.get("current_plan_step", -1)
store_plan = state.get("store_plan", [])
# Check if the step after next exists in the plan
detail_step_index = current_plan_step + 2
if detail_step_index < 0 or detail_step_index >= len(store_plan):
print(f"Warning: Invalid plan step ({detail_step_index}) for KI detailing.")
step_description = "Error: Plan step for KI detailing unavailable."
else:
step_description = store_plan[detail_step_index]
prompt = f"""Using the information provided above, proceed with step {detail_step_index + 1} of your plan to provide the user with NEW and INNOVATIVE Key Issues :
{step_description}"""
return {"messages" : [HumanMessage(content=prompt)]}
##########################################################################
###### CONDITIONAL EDGE FUNCTIONS ######
##########################################################################
def validate_plan(state: State):
"""
Whether to regenerate the plan or to parse it
"""
# Check the last message for "My plan is correct"
if "messages" in state and state["messages"]:
last_message = state["messages"][-1]
if hasattr(last_message, 'content') and "My plan is correct" in last_message.content:
return "parse"
# Default to validate (regenerate) if condition not met or messages are missing
return "validate"
def next_plan_step(state: State, config: ConfigSchema):
"""
Proceed to next plan step (either generate KI or retrieve more data)
"""
current_plan_step = state.get("current_plan_step", -1)
store_plan_len = len(state.get("store_plan", []))
# Simplified logic: go to KI generation if it's the last step based on plan length
if current_plan_step >= store_plan_len - 1:
return "generate_key_issues"
else:
return "detail_step"
def detail_or_data_retriever(state: State, config: ConfigSchema):
"""
Decide whether to detail the query or go straight to data retrieval.
"""
# Check configuration if detailed query is needed
if config["configurable"].get("use_detailed_query"):
# Need to invoke the LLM to get the detailed query
return "chatbot_detail"
else:
# Use the plan step directly as the query
return "data_retriever"
def retrieve_or_process(state: State):
"""
Process the retrieved docs or keep retrieving (based on human_validated flag).
"""
# Check the 'human_validated' flag in the state
# This flag needs to be set externally (e.g., by Streamlit UI or another mechanism)
# before this node is reached after data_retriever.
if state.get('human_validated'):
return "process"
else:
# If not validated, loop back to retrieve more (or wait for validation)
# This assumes data_retriever might be called again or the graph waits.
# In the Streamlit app, the human_validation node allows setting this flag.
return "retrieve"
def build_planner_graph(memory, config):
"""
Builds the planner graph
"""
graph_builder = StateGraph(State)
graph_doc_retriever = build_data_retriever_graph(memory)
graph_doc_processor = build_data_processor_graph(memory)
# Use the generic chatbot node function
graph_builder.add_node("chatbot_planner", lambda state: chatbot_node(state, config))
graph_builder.add_node("validate", validate_node)
# Add node for chatbot interaction when detailed query is needed
graph_builder.add_node("chatbot_detail", lambda state: chatbot_node(state, config))
graph_builder.add_node("parse", parse_plan)
# Pass config to detail_step as it needs it now
graph_builder.add_node("detail_step", lambda state: detail_step(state, config))
graph_builder.add_node("data_retriever", graph_doc_retriever) # Input mapping happens automatically if state keys match
graph_builder.add_node("human_validation", human_validation) # Needs input mapping if HumanValidationState differs significantly
graph_builder.add_node("data_processor", graph_doc_processor) # Needs input mapping if DocProcessorState differs significantly
graph_builder.add_node("concatenate_data", concatenate_data)
# Use the generic chatbot node function
graph_builder.add_node("chatbot_exec_step", lambda state: chatbot_node(state, config))
graph_builder.add_node("generate_ki", generate_ki)
# Use the generic chatbot node function
graph_builder.add_node("chatbot_ki", lambda state: chatbot_node(state, config))
graph_builder.add_node("detail_ki", detail_ki)
# Use the generic chatbot node function
graph_builder.add_node("chatbot_final", lambda state: chatbot_node(state, config))
# Define edges
graph_builder.add_edge("validate", "chatbot_planner")
graph_builder.add_edge("parse", "detail_step")
# Edge from chatbot_detail (after getting detailed query) to data_retriever
graph_builder.add_edge("chatbot_detail", "data_retriever")
graph_builder.add_edge("data_retriever", "human_validation")
graph_builder.add_edge("data_processor", "concatenate_data")
graph_builder.add_edge("concatenate_data", "chatbot_exec_step")
graph_builder.add_edge("generate_ki", "chatbot_ki")
graph_builder.add_edge("chatbot_ki", "detail_ki")
graph_builder.add_edge("detail_ki", "chatbot_final")
graph_builder.add_edge("chatbot_final", "__end__")
# Define conditional edges
graph_builder.add_conditional_edges(
"detail_step",
# Pass config to the conditional function
lambda state: detail_or_data_retriever(state, config),
{"chatbot_detail": "chatbot_detail", "data_retriever": "data_retriever"}
)
graph_builder.add_conditional_edges(
"human_validation",
retrieve_or_process,
# Map 'retrieve' back to 'data_retriever' node, 'process' to 'data_processor'
{"retrieve" : "data_retriever", "process" : "data_processor"}
)
graph_builder.add_conditional_edges(
"chatbot_planner",
validate_plan,
{"parse" : "parse", "validate": "validate"}
)
graph_builder.add_conditional_edges(
"chatbot_exec_step",
# Pass config to the conditional function
lambda state: next_plan_step(state, config),
{"generate_key_issues" : "generate_ki", "detail_step": "detail_step"}
)
# Set entry point
graph_builder.set_entry_point("chatbot_planner")
# Compile the graph
graph = graph_builder.compile(
checkpointer=memory,
# Define interrupt points if needed for human interaction or debugging
interrupt_after=["human_validation", "chatbot_final"],
)
return graph