import os import asyncio import random import re import json import logging from colorama import Fore, Style from groq import AsyncGroq, RateLimitError from mistralai import Mistral from openai import AsyncOpenAI import traceback # Configura logger local logger = logging.getLogger("JadeHeavy") logger.setLevel(logging.INFO) class JadeHeavyAgent: def __init__(self): self.groq_api_key = os.getenv("GROQ_API_KEY") self.mistral_api_key = os.getenv("MISTRAL_API_KEY") self.openrouter_api_key = os.getenv("OPENROUTER_API_KEY") if not self.groq_api_key: logger.warning("GROQ_API_KEY not set. Jade Heavy may fail.") self.groq_client = AsyncGroq(api_key=self.groq_api_key) self.mistral = None if self.mistral_api_key: self.mistral = Mistral(api_key=self.mistral_api_key) else: logger.warning("MISTRAL_API_KEY not set. Mistral model will be skipped or substituted.") self.openrouter = None if self.openrouter_api_key: self.openrouter = AsyncOpenAI( base_url="https://openrouter.ai/api/v1", api_key=self.openrouter_api_key, ) else: logger.warning("OPENROUTER_API_KEY not set. Qwen/OpenRouter models will be skipped.") # Updated Model Map for Generalist Chat self.models = { "Kimi": "moonshotai/kimi-k2-instruct-0905", # Groq (Logic/Reasoning) "Mistral": "mistral-large-latest", # Mistral API "Llama": "openai/gpt-oss-120b", # Groq "Qwen": "qwen/qwen3-coder:free" # OpenRouter (Fallback if key exists) or Groq equivalent # Note: The original script used qwen/qwen3-235b... on OpenRouter. # If no OpenRouter key, we might need a fallback on Groq or skip. } # Judge model (Groq is fast and cheap) self.judge_id = "moonshotai/kimi-k2-instruct-0905" async def _safe_propose(self, model_name, history_text): """Phase 1: Strategic Planning""" # Staggering to avoid rate limits delay_map = {"Kimi": 0, "Mistral": 1.0, "Llama": 2.0, "Qwen": 3.0} await asyncio.sleep(delay_map.get(model_name, 1) + random.uniform(0.1, 0.5)) sys_prompt = ( "You are a Strategic Architect. Create a high-level roadmap to answer the user's request comprehensively.\n" "DO NOT write the final response yet. Just plan the structure and key points.\n" "FORMAT: 1. [INTENT ANALYSIS] 2. [KEY POINTS] 3. [STRUCTURE PROPOSAL]" ) messages = [{"role": "system", "content": sys_prompt}, {"role": "user", "content": history_text}] try: content = "" if model_name == "Mistral" and self.mistral: resp = await self.mistral.chat.complete_async(model=self.models["Mistral"], messages=messages) content = resp.choices[0].message.content elif model_name == "Qwen" and self.openrouter: # Use OpenRouter if available resp = await self.openrouter.chat.completions.create(model="qwen/qwen3-235b-a22b:free", messages=messages) # Using the large free one if possible content = resp.choices[0].message.content else: # Default to Groq (Kimi, Llama, or fallback for others) # If Mistral/OpenRouter key missing, fallback to Llama-3-70b on Groq for diversity? target_model = self.models.get(model_name) if not target_model or (model_name == "Mistral" and not self.mistral) or (model_name == "Qwen" and not self.openrouter): target_model = "openai/gpt-oss-120b" # Fallback resp = await self.groq_client.chat.completions.create( model=target_model, messages=messages, temperature=0.7 ) content = resp.choices[0].message.content if content: return f"--- {model_name} Plan ---\n{content}" except Exception as e: logger.error(f"Error in propose ({model_name}): {e}") return "" return "" async def _safe_expand(self, model_name, history_text, strategy): """Phase 3: Execution/Expansion""" delay_map = {"Kimi": 0, "Mistral": 1.5, "Llama": 3.0, "Qwen": 4.5} await asyncio.sleep(delay_map.get(model_name, 1)) sys_prompt = ( f"You are a Precision Engine. Execute the following plan to answer the user request:\n\n{strategy}\n\n" "Write a detailed, natural, and high-quality response following this plan.\n" "Do not output internal reasoning like '[DECOMPOSITION]', just the final response text." ) messages = [{"role": "system", "content": sys_prompt}, {"role": "user", "content": history_text}] try: content = "" if model_name == "Mistral" and self.mistral: resp = await self.mistral.chat.complete_async(model=self.models["Mistral"], messages=messages) content = resp.choices[0].message.content elif model_name == "Qwen" and self.openrouter: resp = await self.openrouter.chat.completions.create(model="qwen/qwen3-coder:free", messages=messages) content = resp.choices[0].message.content else: target_model = self.models.get(model_name) if not target_model or (model_name == "Mistral" and not self.mistral) or (model_name == "Qwen" and not self.openrouter): target_model = "openai/gpt-oss-120b" resp = await self.groq_client.chat.completions.create( model=target_model, messages=messages, temperature=0.7 ) content = resp.choices[0].message.content if content: return f"[{model_name} Draft]:\n{content}" except Exception as e: logger.error(f"Error in expand ({model_name}): {e}") return "" return "" async def respond(self, history, user_input, user_id="default", vision_context=None): """ Main entry point for the Heavy Agent. History is a list of dicts: [{"role": "user", "content": "..."}...] """ # Prepare context full_context = "" for msg in history[-6:]: # Limit context to last few turns to avoid huge prompts full_context += f"{msg['role'].upper()}: {msg['content']}\n" if vision_context: full_context += f"SYSTEM (Vision): {vision_context}\n" full_context += f"USER: {user_input}\n" agents = ["Kimi", "Mistral", "Llama", "Qwen"] # --- PHASE 1: STRATEGY --- logger.info("Jade Heavy: Phase 1 - Planning...") tasks = [self._safe_propose(m, full_context) for m in agents] results = await asyncio.gather(*tasks) valid_strats = [s for s in results if s] if not valid_strats: return "Failed to generate a plan.", None, history # --- PHASE 2: PRUNING (Select Best Plan) --- logger.info("Jade Heavy: Phase 2 - Pruning...") prune_prompt = ( f"User Request Context:\n{full_context}\n\nProposed Plans:\n" + "\n".join(valid_strats) + "\n\nTASK: SELECT THE SINGLE MOST ROBUST AND HELPFUL PLAN. Return ONLY the content of the best plan." ) try: best_strat_resp = await self.groq_client.chat.completions.create( model=self.judge_id, messages=[{"role":"user","content":prune_prompt}], temperature=0.5 ) best_strat = best_strat_resp.choices[0].message.content except Exception as e: logger.error(f"Pruning failed: {e}") best_strat = valid_strats[0] # Fallback to first plan # --- PHASE 3: EXPANSION (Drafting Responses) --- logger.info("Jade Heavy: Phase 3 - Expansion...") tasks_exp = [self._safe_expand(m, full_context, best_strat) for m in agents] results_exp = await asyncio.gather(*tasks_exp) valid_sols = [s for s in results_exp if s] if not valid_sols: return "Failed to generate drafts.", None, history # --- PHASE 4: VERDICT (Synthesis) --- logger.info("Jade Heavy: Phase 4 - Verdict...") council_prompt = ( f"User Request:\n{full_context}\n\nCandidate Responses:\n" + "\n".join(valid_sols) + "\n\nTASK: Synthesize the best parts of these drafts into a FINAL, PERFECT RESPONSE." "The response should be natural, helpful, and high-quality. Do not mention the agents or the process." ) final_answer = "" try: resp = await self.groq_client.chat.completions.create( model=self.judge_id, messages=[{"role":"system","content":"You are the Chief Editor."},{"role":"user","content":council_prompt}], temperature=0.5 ) final_answer = resp.choices[0].message.content except Exception as e: logger.error(f"Verdict failed: {e}") final_answer = valid_sols[0].replace(f"[{agents[0]} Draft]:\n", "") # Fallback # Update History history.append({"role": "user", "content": user_input}) history.append({"role": "assistant", "content": final_answer}) # Audio (Optional/Placeholder - returning None for now or use TTS if needed) # The user said "backend focuses on request", so we can skip TTS generation here # or implement it if JadeAgent does it. The original code uses `jade_agent.tts`. # For Heavy mode, maybe we skip audio for speed, or add it later. # I'll return None for audio path. return final_answer, None, history