flare / chat_handler.py
ciyidogan's picture
Update chat_handler.py
ec93388 verified
raw
history blame
17 kB
"""
Flare – Chat Handler (Refactored with LLM Factory)
==========================================
"""
import re, json, sys, httpx, os
from datetime import datetime
from typing import Dict, List, Optional, Any
from fastapi import APIRouter, HTTPException, Header
from pydantic import BaseModel
import requests
from prompt_builder import build_intent_prompt, build_parameter_prompt
from utils import log
from api_executor import call_api as execute_api
from config_provider import ConfigProvider
from validation_engine import validate
from session import session_store, Session
# Initialize router
router = APIRouter()
# ───────────────────────── GLOBAL VARS ───────────────────────── #
cfg = ConfigProvider.get()
llm_provider = None
# ───────────────────────── HELPERS ───────────────────────── #
def _trim_response(raw: str) -> str:
"""
Remove everything after the first logical assistant block or intent tag.
Also strips trailing 'assistant' artifacts and prompt injections.
"""
# Stop at our own rules if model leaked them
for stop in ["#DETECTED_INTENT", "⚠️", "\nassistant", "assistant\n", "assistant"]:
idx = raw.find(stop)
if idx != -1:
raw = raw[:idx]
# Normalise selamlama
raw = re.sub(r"Hoş[\s-]?geldin(iz)?", "Hoş geldiniz", raw, flags=re.IGNORECASE)
return raw.strip()
def _safe_intent_parse(raw: str) -> tuple[str, str]:
"""Extract intent name and extra tail."""
m = re.search(r"#DETECTED_INTENT:\s*([A-Za-z0-9_-]+)", raw)
if not m:
return "", raw
name = m.group(1)
# Remove 'assistant' suffix if exists
if name.endswith("assistant"):
name = name[:-9] # Remove last 9 chars ("assistant")
log(f"πŸ”§ Removed 'assistant' suffix from intent name")
tail = raw[m.end():]
log(f"🎯 Parsed intent: {name}")
return name, tail
# ───────────────────────── LLM SETUP ───────────────────────── #
def setup_llm_provider():
"""Initialize LLM provider using factory pattern"""
global llm_provider
try:
from llm_factory import LLMFactory
llm_provider = LLMFactory.create_provider()
log("βœ… LLM provider initialized successfully")
except Exception as e:
log(f"❌ Failed to initialize LLM provider: {e}")
raise
# ───────────────────────── LLM GENERATION ───────────────────────── #
async def llm_generate(s: Session, prompt: str, user_msg: str) -> str:
"""Call LLM provider with proper error handling"""
global llm_provider
if llm_provider is None:
setup_llm_provider()
try:
# Get version config from session
version = s.get_version_config()
if not version:
# Fallback: get from project config
project = next((p for p in cfg.projects if p.name == s.project_name), None)
if not project:
raise ValueError(f"Project not found: {s.project_name}")
version = next((v for v in project.versions if v.published), None)
if not version:
raise ValueError("No published version found")
log(f"πŸš€ Calling LLM for session {s.session_id[:8]}...")
log(f"πŸ“‹ Prompt preview (first 200 chars): {prompt[:200]}...")
# Call the configured LLM provider
raw = await llm_provider.generate(
user_input=user_msg,
system_prompt=prompt,
context=s.chat_history[-10:] if s.chat_history else []
)
log(f"πŸͺ„ LLM raw response: {raw[:100]}...")
return raw
except requests.exceptions.Timeout:
log(f"⏱️ LLM timeout for session {s.session_id[:8]}")
raise HTTPException(status_code=504, detail="LLM request timed out")
except Exception as e:
log(f"❌ LLM error: {e}")
raise HTTPException(status_code=500, detail=f"LLM error: {str(e)}")
# ───────────────────────── PARAMETER EXTRACTION ───────────────────────── #
def _extract_parameters_from_response(raw: str, session: Session, intent_config) -> bool:
"""Extract parameters from the LLM response"""
try:
# Look for JSON block in response
json_match = re.search(r'```json\s*(.*?)\s*```', raw, re.DOTALL)
if not json_match:
# Try to find JSON without code block
json_match = re.search(r'\{[^}]+\}', raw)
if not json_match:
log("❌ No JSON found in response")
return False
json_str = json_match.group(1) if '```' in raw else json_match.group(0)
params = json.loads(json_str)
any_valid = False
for param_name, param_value in params.items():
# Find parameter config
param_config = next(
(p for p in intent_config.parameters if p.name == param_name),
None
)
if not param_config:
log(f"⚠️ Parameter config not found for: {param_name}")
continue
# Validate parameter
if validate(str(param_value), param_config):
session.variables[param_config.variable_name] = str(param_value)
any_valid = True
log(f"βœ… Extracted {param_name}={param_value} β†’ {param_config.variable_name}")
else:
log(f"❌ Invalid {param_name}={param_value}")
return any_valid
except json.JSONDecodeError as e:
log(f"❌ JSON parsing error: {e}")
log(f"❌ Failed to parse: {raw[:200]}")
return False
except Exception as e:
log(f"❌ Parameter processing error: {e}")
return False
# ───────────────────────── API EXECUTION ───────────────────────── #
async def _execute_api_call(session: Session, intent_config) -> str:
"""Execute API call and return humanized response"""
try:
session.state = "call_api"
api_name = intent_config.action
api_config = cfg.get_api(api_name)
if not api_config:
log(f"❌ API config not found: {api_name}")
session.reset_flow()
return intent_config.fallback_error_prompt or "İşlem başarısız oldu."
log(f"πŸ“‘ Calling API: {api_name}")
log(f"πŸ“¦ API variables: {session.variables}")
# Execute API call with session
response = execute_api(api_config, session)
api_json = response.json()
log(f"βœ… API response: {api_json}")
# Humanize response
session.state = "humanize"
if api_config.response_prompt:
prompt = api_config.response_prompt.replace(
"{{api_response}}",
json.dumps(api_json, ensure_ascii=False)
)
human_response = await llm_generate(session, prompt, json.dumps(api_json))
session.reset_flow()
return human_response if human_response else f"İşlem sonucu: {api_json}"
else:
session.reset_flow()
return f"İşlem tamamlandı: {api_json}"
except requests.exceptions.Timeout:
log(f"⏱️ API timeout: {api_name}")
session.reset_flow()
return intent_config.fallback_timeout_prompt or "İşlem zaman aşımına uğradı."
except Exception as e:
log(f"❌ API call error: {e}")
session.reset_flow()
return intent_config.fallback_error_prompt or "İşlem sırasında bir hata oluştu."
# ───────────────────────── REQUEST MODELS ───────────────────────── #
class SessionRequest(BaseModel):
project_name: str
test_mode: bool = False
class ChatRequest(BaseModel):
message: str
class StartRequest(BaseModel):
project_name: str
version_number: Optional[int] = None # Opsiyonel, belirtilmezse published olan en bΓΌyΓΌk version no'yu kullan
# ───────────────────────── API ENDPOINTS ───────────────────────── #
@router.post("/start_session", response_model=ChatResponse)
async def start_session(req: StartRequest):
"""Create new session"""
try:
# Validate project exists
project = next((p for p in cfg.projects if p.name == req.project_name and p.enabled), None)
if not project:
raise HTTPException(404, f"Project '{req.project_name}' not found or disabled")
# Find version
if req.version_number:
# Specific version requested
version = next((v for v in project.versions if v.id == req.version_number), None)
if not version:
raise HTTPException(404, f"Version {req.version_number} not found for project '{req.project_name}'")
else:
# Find published version with highest version number
published_versions = [v for v in project.versions if v.published]
if not published_versions:
raise HTTPException(404, f"No published version for project '{req.project_name}'")
# Sort by version number (id) and get the highest
version = max(published_versions, key=lambda v: v.id)
# Create session with version config
session = session_store.create_session(req.project_name, version)
greeting = "Hoş geldiniz! Size nasıl yardımcı olabilirim?"
session.add_turn("assistant", greeting)
log(f"βœ… Session created for project '{req.project_name}' version {version.id} (highest published)")
return ChatResponse(session_id=session.session_id, answer=greeting)
except HTTPException:
raise
except Exception as e:
log(f"❌ Session creation error: {e}")
raise HTTPException(500, f"Session creation failed: {str(e)}")
@router.post("/chat")
async def chat(req: ChatRequest, x_session_id: str = Header(...)):
"""Process chat message"""
try:
# Get session
session = session_store.get(x_session_id)
if not session:
raise HTTPException(status_code=404, detail="Session not found or expired")
# Add user message to history
session.add_message("user", req.message)
log(f"πŸ’¬ User [{session.session_id[:8]}...]: {req.message}")
# Get project and version config
project = next((p for p in cfg.projects if p.name == session.project_name), None)
if not project:
raise HTTPException(status_code=404, detail=f"Project '{session.project_name}' not found")
version = session.get_version_config()
if not version:
raise HTTPException(status_code=400, detail="Version config not found in session")
# Process based on current state
if session.state == "idle":
# Build intent detection prompt
prompt = build_intent_prompt(version, session.chat_history, project.default_locale)
raw = await llm_generate(session, prompt, req.message)
# Check for intent
intent_name, tail = _safe_intent_parse(raw)
if intent_name:
# Find intent config
intent_config = next((i for i in version.intents if i.name == intent_name), None)
if intent_config:
session.current_intent = intent_name
session.intent_config = intent_config
session.state = "collect_params"
log(f"🎯 Intent detected: {intent_name}")
# Check if parameters were already extracted
if tail and _extract_parameters_from_response(tail, session, intent_config):
log("πŸ“¦ Some parameters extracted from initial response")
# Check what parameters are missing
missing_params = [
p for p in intent_config.parameters
if p.required and p.variable_name not in session.variables
]
if not missing_params:
# All required parameters collected, execute API
response = await _execute_api_call(session, intent_config)
session.add_message("assistant", response)
return {"response": response, "intent": intent_name, "state": "completed"}
else:
# Need to collect more parameters
param_prompt = build_parameter_prompt(
intent_config,
session.variables,
session.chat_history,
project.default_locale
)
param_question = await llm_generate(session, param_prompt, req.message)
clean_question = _trim_response(param_question)
session.add_message("assistant", clean_question)
return {"response": clean_question, "intent": intent_name, "state": "collecting_params"}
else:
log(f"⚠️ Unknown intent: {intent_name}")
# No intent detected, return general response
clean_response = _trim_response(raw)
session.add_message("assistant", clean_response)
return {"response": clean_response, "state": "idle"}
elif session.state == "collect_params":
# Continue parameter collection
intent_config = session.intent_config
# Try to extract parameters from user message
param_prompt = f"""
Extract parameters from user message: "{req.message}"
Expected parameters:
{json.dumps([{
'name': p.name,
'type': p.type,
'required': p.required,
'extraction_prompt': p.extraction_prompt
} for p in intent_config.parameters if p.variable_name not in session.variables], ensure_ascii=False)}
Return as JSON object with parameter names as keys.
"""
raw = await llm_generate(session, param_prompt, req.message)
_extract_parameters_from_response(raw, session, intent_config)
# Check what parameters are still missing
missing_params = [
p for p in intent_config.parameters
if p.required and p.variable_name not in session.variables
]
if not missing_params:
# All parameters collected, execute API
response = await _execute_api_call(session, intent_config)
session.add_message("assistant", response)
return {"response": response, "intent": session.current_intent, "state": "completed"}
else:
# Still need more parameters
param_prompt = build_parameter_prompt(
intent_config,
session.variables,
session.chat_history,
project.default_locale
)
param_question = await llm_generate(session, param_prompt, req.message)
clean_question = _trim_response(param_question)
session.add_message("assistant", clean_question)
return {"response": clean_question, "intent": session.current_intent, "state": "collecting_params"}
else:
# Unknown state, reset
session.reset_flow()
return {"response": "Bir hata oluştu, lütfen tekrar deneyin.", "state": "error"}
except HTTPException:
raise
except Exception as e:
log(f"❌ Chat error: {e}")
import traceback
traceback.print_exc()
raise HTTPException(status_code=500, detail=str(e))
# Initialize LLM on module load
setup_llm_provider()