api.py

import logging
import time
import uvicorn
import requests
import os # Added import for environment variables

from fastapi import FastAPI, HTTPException, Body
from pydantic import BaseModel
from contextlib import asynccontextmanager
from typing import List, Dict, Any

# Import necessary components from your kig_core library
# Ensure kig_core is in the Python path or installed as a package
try:
    from kig_core.config import settings # Loads config on import
    from kig_core.schemas import PlannerState, KeyIssue as KigKeyIssue, GraphConfig
    from kig_core.planner import build_graph
    from kig_core.graph_client import neo4j_client # Import the initialized client instance
    from langchain_core.messages import HumanMessage
except ImportError as e:
    print(f"Error importing kig_core components: {e}")
    print("Please ensure kig_core is in your Python path or installed.")
    # You might want to exit or raise a clearer error if imports fail
    raise


import google.generativeai as genai

# Configure logging for the API
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

# --- Pydantic Models for API Request/Response ---

class KeyIssueRequest(BaseModel):
    """Request body containing the user's technical query."""
    query: str

class KeyIssueResponse(BaseModel):
    """Response body containing the generated key issues."""
    key_issues: List[KigKeyIssue] # Use the KeyIssue schema from kig_core

class SpecificityEvaluationRequest(BaseModel):
    title: str
    description: str
    technical_topic: str

class SpecificityScore(BaseModel):
    predicted_class: str
    score: float

class SpecificityEvaluationResponse(BaseModel):
    problematic: str
    specificity: SpecificityScore

class ProblemDescriptionRequest(BaseModel):
    description: str
    technical_topic: str

class ProblemDescriptionResponse(BaseModel):
    problem_description: str

# --- Global Variables / State ---
# Keep the graph instance global for efficiency if desired,
# but consider potential concurrency issues if graph/LLMs have state.
# Rebuilding on each request is safer for statelessness.
app_graph = None # Will be initialized at startup
gemini_client = None # Will be initialized at startup

# --- Application Lifecycle (Startup/Shutdown) ---
@asynccontextmanager
async def lifespan(app: FastAPI):
    """Handles startup and shutdown events."""
    global app_graph
    logger.info("API starting up...")
    # Initialize Neo4j client (already done on import by graph_client.py)
    # Verify connection (optional, already done by graph_client on init)
    try:
        logger.info("Verifying Neo4j connection...")
        neo4j_client._get_driver().verify_connectivity()
        logger.info("Neo4j connection verified.")
    except Exception as e:
        logger.error(f"Neo4j connection verification failed on startup: {e}", exc_info=True)
        # Decide if the app should fail to start
        # raise RuntimeError("Failed to connect to Neo4j on startup.") from e

    # Build the LangGraph application
    logger.info("Building LangGraph application...")
    try:
        app_graph = build_graph()
        logger.info("LangGraph application built successfully.")
    except Exception as e:
        logger.error(f"Failed to build LangGraph application on startup: {e}", exc_info=True)
        # Decide if the app should fail to start
        raise RuntimeError("Failed to build LangGraph on startup.") from e

    # Initialize Gemini Client
    logger.info("Initializing Gemini client...")
    if genai:
        try:
            # Assuming GEMINI_API_KEY is set in environment or loaded via settings
            api_key = os.getenv("GEMINI_API_KEY") or getattr(settings, "GEMINI_API_KEY", None)
            if not api_key:
                 raise ValueError("GEMINI_API_KEY not found in environment or settings.")
            genai.configure(api_key=api_key)
            logger.info("Gemini client configured successfully.")
        except Exception as e:
            logger.error(f"Failed to configure Gemini client: {e}", exc_info=True)
    else:
        logger.warning("Gemini library not imported. Endpoints requiring Gemini will not work.")

    yield # API runs here

    # --- Shutdown ---
    logger.info("API shutting down...")
    # Close Neo4j connection (handled by atexit in graph_client.py)
    # neo4j_client.close() # Usually not needed due to atexit registration
    logger.info("Neo4j client closed (likely via atexit).")
    logger.info("API shutdown complete.")


# --- FastAPI Application ---
app = FastAPI(
    title="Key Issue Generator, Specificity, and Description API",
    description="API to generate Key Issues, evaluate problematic specificity, and generate problem descriptions.",
    version="1.1.0",
    lifespan=lifespan # Use the lifespan context manager
)

# --- API Endpoint ---
# API state check route
@app.get("/")
def read_root():
    return {"status": "ok"}

@app.post("/generate-key-issues", response_model=KeyIssueResponse)
async def generate_issues(request: KeyIssueRequest):
    """
    Accepts a technical query and returns a list of generated Key Issues.
    """
    global app_graph
    if app_graph is None:
        logger.error("Graph application is not initialized.")
        raise HTTPException(status_code=503, detail="Service Unavailable: Graph not initialized")

    user_query = request.query
    if not user_query:
        raise HTTPException(status_code=400, detail="Query cannot be empty.")

    logger.info(f"Received request to generate key issues for query: '{user_query[:100]}...'")
    start_time = time.time()

    try:
        # --- Prepare Initial State for LangGraph ---
        # Note: Ensure PlannerState aligns with what build_graph expects
        initial_state: PlannerState = {
            "user_query": user_query,
            "messages": [HumanMessage(content=user_query)],
            "plan": [],
            "current_plan_step_index": -1, # Or as expected by your graph's entry point
            "step_outputs": {},
            "key_issues": [],
            "error": None
        }

        # --- Define Configuration (e.g., Thread ID for Memory) ---
        # Using a simple thread ID; adapt if using persistent memory
        # import hashlib
        # thread_id = hashlib.sha256(user_query.encode()).hexdigest()[:8]
        # config: GraphConfig = {"configurable": {"thread_id": thread_id}}
        # If not using memory, config can be simpler or empty based on LangGraph version
        config: GraphConfig = {"configurable": {}} # Adjust if thread_id/memory is needed

        # --- Execute the LangGraph Workflow ---
        logger.info("Invoking LangGraph workflow...")
        # Use invoke for a single result, or stream if you need intermediate steps
        final_state = await app_graph.ainvoke(initial_state, config=config)
        # If using stream:
        # final_state = None
        # async for step_state in app_graph.astream(initial_state, config=config):
        #     # Process intermediate states if needed
        #     node_name = list(step_state.keys())[0]
        #     logger.debug(f"Graph step completed: {node_name}")
        #     final_state = step_state[node_name] # Get the latest full state output

        end_time = time.time()
        logger.info(f"Workflow finished in {end_time - start_time:.2f} seconds.")

        # --- Process Final Results ---
        if final_state is None:
             logger.error("Workflow execution did not produce a final state.")
             raise HTTPException(status_code=500, detail="Workflow execution failed to produce a result.")

        if final_state.get("error"):
            error_msg = final_state.get("error", "Unknown error")
            logger.error(f"Workflow failed with error: {error_msg}")
            # Map internal errors to appropriate HTTP status codes
            status_code = 500 # Internal Server Error by default
            if "Neo4j" in error_msg or "connection" in error_msg.lower():
                status_code = 503 # Service Unavailable (database issue)
            elif "LLM error" in error_msg or "parse" in error_msg.lower():
                 status_code = 502 # Bad Gateway (issue with upstream LLM)

            raise HTTPException(status_code=status_code, detail=f"Workflow failed: {error_msg}")

        # --- Extract Key Issues ---
        # Ensure the structure matches KeyIssueResponse and KigKeyIssue Pydantic model
        generated_issues_data = final_state.get("key_issues", [])

        # Validate and convert if necessary (Pydantic usually handles this via response_model)
        try:
            # Pydantic will validate against KeyIssueResponse -> List[KigKeyIssue]
            response_data = {"key_issues": generated_issues_data}
            logger.info(f"Successfully generated {len(generated_issues_data)} key issues.")
            return response_data
        except Exception as pydantic_error: # Catch potential validation errors
             logger.error(f"Failed to validate final key issues against response model: {pydantic_error}", exc_info=True)
             logger.error(f"Data that failed validation: {generated_issues_data}")
             raise HTTPException(status_code=500, detail="Internal error: Failed to format key issues response.")


    except HTTPException as http_exc:
        # Re-raise HTTPExceptions directly
        raise http_exc
    except ConnectionError as e:
        logger.error(f"Connection Error during API request: {e}", exc_info=True)
        raise HTTPException(status_code=503, detail=f"Service Unavailable: {e}")
    except ValueError as e:
         logger.error(f"Value Error during API request: {e}", exc_info=True)
         raise HTTPException(status_code=400, detail=f"Bad Request: {e}") # Often input validation issues
    except Exception as e:
        logger.error(f"An unexpected error occurred during API request: {e}", exc_info=True)
        raise HTTPException(status_code=500, detail=f"Internal Server Error: An unexpected error occurred.")


@app.post("/evaluate-specificity", response_model=SpecificityEvaluationResponse)
async def evaluation(request: SpecificityEvaluationRequest):
    """
    Generates a technical problematic using Gemini based on title, description,
    and topic, then evaluates its specificity using an external API (fine-tuned model for specificity).
    """
    # Check if Gemini library was imported and configured
    if not genai:
        logger.error("Gemini client is not available or configured.")
        raise HTTPException(status_code=503, detail="Service Unavailable: Gemini client not configured")

    title = request.title
    description = request.description
    technical_topic = request.technical_topic

    if not all([title, description, technical_topic]):
        raise HTTPException(status_code=400, detail="Missing required fields: title, description, or technical_topic.")

    logger.info("Received request for specificity evaluation.")
    logger.debug(f"Title: {title}, Topic: {technical_topic}") # Avoid logging full description unless needed

    # --- 1. Generate Problematic using Gemini ---
    prompt = f"""I want you to create a technical problematic using a key issue composed of a title and a detailed description.
    Here is the title of the key issue to deal with: <title>{title}</title>
    
    And here is the associated description: <description>{description}</description>
    
    This key issue is part of the following technical topic: <technical_topic>{technical_topic}</technical_topic>
    
    The problematic must be in interrogative form.
    As the output, I only want you to provide the problematic found, nothing else.
    
    Here are examples of problematics that you could create, it shows the level of specificity that you should aim for:
    
    Example 1: 'How can a method for allocating radio resources in a non-GSO satellite communication system, operating in frequency bands shared with geostationary satellite systems and particularly in high frequency bands such as Ka, minimize interference to geostationary systems, without causing reduced geographic coverage due to fixed high separation angle thresholds or incurring cost and sub-optimality from over-dimensioning the non-GSO satellite constellation?'
    Example 2: 'How to address the vulnerability of on-aircraft avionics software update integrity checks to system compromises and the logistical challenges of cryptographic key management in digital signature solutions, in order to establish a more secure and logistically efficient method for updating safety-critical avionics equipment?'
    Example 3: 'How can SIM cards be protected against collision attacks that aim to retrieve the secret key Ki by analyzing the input and output of the authentication algorithm during the standard GSM authentication process, given that current tamper-proof measures are insufficient to prevent this type of key extraction?'
    Example 4: 'How can a Trusted Application in a GlobalPlatform compliant TEE overcome the GP specification limitations that enforce client blocking during task execution, prevent partial task execution, and delete TA execution context between commands, to function as a persistent server with stateful sessions and asynchronous communication capabilities, thereby enabling server functionalities like continuous listening and non-blocking send/receive, currently impossible due to GP's sequential task processing and stateless TA operation?'
    
    As far as possible, avoid using acronyms in the problematic.
    Try to be about the same length as the examples if possible."""

    try:
        logger.info("Calling Gemini API to generate problematic...")
        # Use the specified model and configuration
        model_name = "gemini-2.5-flash-preview-04-17"
        model = genai.GenerativeModel(model_name)

        response = model.generate_content(prompt)
        # Extract the result
        problematic_result = response.text.strip()
        logger.info("Successfully generated problematic from Gemini.")
        logger.debug(f"Generated problematic: {problematic_result[:200]}...") # Log snippet

    except Exception as e:
        logger.error(f"Error calling Gemini API: {e}", exc_info=True)
        # Check for specific Gemini API errors if possible
        raise HTTPException(status_code=502, detail=f"Failed to generate problematic using LLM: {e}")

    if not problematic_result:
         logger.error("Gemini API returned an empty result.")
         raise HTTPException(status_code=502, detail="LLM returned an empty problematic.")

    # --- 2. Evaluate Specificity using External API ---
    API_URL = "https://organizedprogrammers-fastapi-problematic-specificity.hf.space"
    endpoint = f"{API_URL}/predict"
    data = {"text": problematic_result}

    try:
        logger.info(f"Calling specificity prediction API at {endpoint}...")
        prediction_response = requests.post(endpoint, json=data, timeout=30) # Added timeout
        prediction_response.raise_for_status() # Raise HTTPError for bad responses (4xx or 5xx)

        score_data = prediction_response.json()
        logger.info(f"Successfully received specificity score: {score_data}")

        # Validate the received score data against Pydantic model
        try:
            specificity_score = SpecificityScore(**score_data)
        except Exception as pydantic_error: # Catch validation errors
            logger.error(f"Failed to validate specificity score response: {pydantic_error}", exc_info=True)
            logger.error(f"Invalid data received from specificity API: {score_data}")
            raise HTTPException(status_code=502, detail="Invalid response format from specificity prediction API.")

    except requests.exceptions.RequestException as e:
        logger.error(f"Error calling specificity prediction API: {e}", exc_info=True)
        raise HTTPException(status_code=502, detail=f"Failed to call specificity prediction API: {e}")
    except Exception as e: # Catch other potential errors like JSON decoding
         logger.error(f"Unexpected error during specificity evaluation: {e}", exc_info=True)
         raise HTTPException(status_code=500, detail=f"Internal error during specificity evaluation: {e}")


    # --- 3. Return Combined Result ---
    final_response = SpecificityEvaluationResponse(
        problematic=problematic_result,
        specificity=specificity_score
    )
    return final_response


@app.post("/create-problem-description", response_model=ProblemDescriptionResponse)
async def gen_problem_description(request: ProblemDescriptionRequest):
    """
    Generates a problem description using Gemini based on a detailed description
    and technical topic.
    """
    description = request.description
    technical_topic = request.technical_topic

    if not description or not technical_topic:
        raise HTTPException(status_code=400, detail="Missing required fields: description or technical_topic.")

    logger.info("Received request for problem description generation.")
    logger.debug(f"Topic: {technical_topic}") # Avoid logging full description unless necessary

    # --- Construct Prompt ---
    prompt = f"""I want you to create a problem description using a key issue explained in a detailed description.

    Here is the description of the key issue: <description>{description}</description>
    
    This key issue is part of the following technical topic: <technical_topic>{technical_topic}</technical_topic>
    
    As the output, I only want you to provide the problem description found, nothing else.
    
    Here are examples of problem descriptions that you could create, it shows the level of specificity that you should aim for:
    
    Example 1: 'I am working on enhancing security in 4G and 5G telecommunications networks, specifically in the area of network slicing. My goal is to address vulnerabilities in the isolation of slices and improve the integrity of data transmission across different network slices, ensuring they are properly protected against unauthorized access and attacks.'
    Example 2: 'I am working on improving user authentication in the context of 3GPP (3rd Generation Partnership Project) networks. Specifically, I need to solve issues related to the security and efficiency of the authentication process in the 5G network architecture, focusing on the use of cryptographic keys and identity management.'
    
    As far as possible, avoid using acronyms in the problem description.
    Try to be about the same length as the examples if possible."""

    try:
        logger.info("Calling Gemini API to generate problem description...")
        # Use the specified model and configuration
        model_name = "gemini-2.5-flash-preview-04-17"
        model = genai.GenerativeModel(model_name)

        response = model.generate_content(prompt)
        # Extract the result
        result_text = response.text.strip()
        logger.info("Successfully generated problem description from Gemini.")
        logger.debug(f"Generated description: {result_text[:200]}...") # Log snippet

    except Exception as e:
        logger.error(f"Error calling Gemini API for problem description: {e}", exc_info=True)
        # Check for specific Gemini API errors if possible
        raise HTTPException(status_code=502, detail=f"Failed to generate problem description using LLM: {e}")

    if not result_text:
         logger.error("Gemini API returned an empty result for problem description.")
         raise HTTPException(status_code=502, detail="LLM returned an empty problem description.")

    # --- Return Result ---
    # Return as JSON using the Pydantic model
    return ProblemDescriptionResponse(problem_description=result_text)

    # Alternative: Return as plain text
    # return PlainTextResponse(content=result_text)

# --- How to Run ---
if __name__ == "__main__":
    # Make sure to set environment variables for config (NEO4J_URI, NEO4J_PASSWORD, GEMINI_API_KEY, etc.)
    # or have a .env file in the same directory where you run this script.
    print("Starting API server...")
    print("Ensure required environment variables (e.g., NEO4J_URI, NEO4J_PASSWORD, GEMINI_API_KEY) are set or .env file is present.")
    # Run with uvicorn: uvicorn api:app --reload --host 0.0.0.0 --port 8000
    # The --reload flag is good for development. Remove it for production.
    uvicorn.run("api:app", host="0.0.0.0", port=8000, reload=True) # Use reload=False for production