This commit is contained in:
xavier committed
1 parent
83e70eadf6
commit
134312ebff
4 files changed
+290
-59
No files matched your search
+29
-5
@@ -8,6 +8,16 @@ from sqlalchemy.exc import IntegrityError
|
||||
from .models import SystemSettings, AIModel, FinancialLog
|
||||
from datetime import datetime, timezone
|
||||
|
||||
# --- LOGGER GLOBAL POUR LE TERMINAL ---
|
||||
activity_logs = []
|
||||
|
||||
def log_activity(msg: str):
|
||||
ts = datetime.now(timezone.utc).strftime('%H:%M:%S')
|
||||
activity_logs.append(f"[{ts}] {msg}")
|
||||
if len(activity_logs) > 100:
|
||||
activity_logs.pop(0)
|
||||
# --------------------------------------
|
||||
|
||||
def determine_domain(model_id: str) -> str:
|
||||
mid = model_id.lower()
|
||||
if "vision" in mid or "vl" in mid or "omni" in mid: return "Vision & Texte"
|
||||
@@ -74,6 +84,7 @@ def update_finance_db(db, provider, balance, usage):
|
||||
except Exception as e: print(f"Finance DB Error: {e}")
|
||||
|
||||
async def sync_providers_models(db: Session, settings: SystemSettings, sync_type: str = "Automatique"):
|
||||
log_activity(f"Lancement de la synchronisation des modèles ({sync_type})...")
|
||||
added = 0
|
||||
models_to_process = {}
|
||||
async with httpx.AsyncClient(timeout=90.0) as client:
|
||||
@@ -171,6 +182,7 @@ async def sync_providers_models(db: Session, settings: SystemSettings, sync_type
|
||||
db.rollback()
|
||||
|
||||
await sync_finances(db, settings)
|
||||
log_activity(f"Synchronisation terminée : {added} modèles analysés.")
|
||||
return {"status": "success", "models_processed": added}
|
||||
|
||||
def get_client_for_model(db: Session, model_id: str, settings: SystemSettings):
|
||||
@@ -194,7 +206,6 @@ async def ask_agent(client, model_id, messages, provider="openrouter"):
|
||||
kwargs = {"model": model_id, "messages": messages}
|
||||
if provider == "openrouter":
|
||||
kwargs["extra_headers"] = {"HTTP-Referer": "https://aethas38.duckdns.org", "X-Title": "AETHAS38 Orchestrator"}
|
||||
# Activation du plugin natif de compression d'OpenRouter pour éviter le dépassement de contexte
|
||||
kwargs["extra_body"] = {"plugins": [{"id": "context-compression"}]}
|
||||
|
||||
resp = await client.chat.completions.create(**kwargs)
|
||||
@@ -211,70 +222,80 @@ async def run_orchestrator(db: Session, history: list, settings: SystemSettings,
|
||||
p_mod = config.get("prompter")
|
||||
if not p_mod: p_mod = "gemini-3.5-flash-lite"
|
||||
|
||||
log_activity(f"Démarrage de l'orchestration. Modèle Prompteur: {p_mod}")
|
||||
p_client, p_prov = get_client_for_model(db, p_mod, settings)
|
||||
|
||||
# --- WORKFLOW MAP-REDUCE : PRÉ-TRAITEMENT SÉQUENTIEL & CHUNKING ---
|
||||
files_context = ""
|
||||
if extracted_files:
|
||||
log_activity(f"Traitement Map-Reduce de {len(extracted_files)} fichier(s) joint(s).")
|
||||
async def process_single_file(f):
|
||||
file_sys = "You are an expert data analyst and senior developer. Extract the most important technical information from the file without losing critical code syntax."
|
||||
content = f['content']
|
||||
chunk_size = 150000 # Environ 35k à 40k tokens par morceau pour rester très large par rapport aux limites
|
||||
chunk_size = 150000
|
||||
|
||||
# CHUNKING : Découpage intelligent si le fichier est massif
|
||||
if len(content) > chunk_size:
|
||||
chunks = [content[i:i+chunk_size] for i in range(0, len(content), chunk_size)]
|
||||
chunk_analyses = []
|
||||
log_activity(f"Fichier lourd ({f['name']}): Chunking en {len(chunks)} morceaux.")
|
||||
for idx, chunk in enumerate(chunks):
|
||||
file_prompt = f"Demande de l'utilisateur : '{original_user_text}'.\n\nPartie {idx+1}/{len(chunks)} du fichier '{f['name']}'. Analysez, extrayez et résumez le code, VBA, SQL ou les données pertinentes.\n\nContenu :\n```\n{chunk}\n```"
|
||||
try:
|
||||
analysis = await ask_agent(p_client, p_mod, [{"role": "system", "content": file_sys}, {"role": "user", "content": file_prompt}], p_prov)
|
||||
chunk_analyses.append(analysis)
|
||||
log_activity(f"Analyse chunk {idx+1}/{len(chunks)} pour {f['name']} réussie.")
|
||||
except Exception as e:
|
||||
chunk_analyses.append(f"[Erreur sur la partie {idx+1}: {str(e)}]")
|
||||
log_activity(f"Erreur chunk {idx+1}/{len(chunks)} pour {f['name']}: {str(e)}")
|
||||
|
||||
await asyncio.sleep(1.5) # Pause anti-spam (429) entre les morceaux
|
||||
await asyncio.sleep(1.5)
|
||||
|
||||
return f"\n\n--- Extraction du fichier {f['name']} (en {len(chunks)} parties) ---\n" + "\n".join(chunk_analyses)
|
||||
else:
|
||||
file_prompt = f"Demande de l'utilisateur : '{original_user_text}'.\n\nAnalysez le fichier ci-dessous. Extrayez, résumez et conservez méticuleusement tout le code, les macros VBA, les requêtes SQL, ou les données métier pertinentes pour répondre à la demande.\n\nFichier : {f['name']}\nContenu :\n```\n{content}\n```"
|
||||
try:
|
||||
analysis = await ask_agent(p_client, p_mod, [{"role": "system", "content": file_sys}, {"role": "user", "content": file_prompt}], p_prov)
|
||||
log_activity(f"Analyse intégrale de {f['name']} réussie.")
|
||||
return f"\n\n--- Extraction du fichier {f['name']} ---\n{analysis}"
|
||||
except Exception as e:
|
||||
log_activity(f"Erreur d'analyse sur {f['name']}: {str(e)}")
|
||||
return f"\n\n--- Erreur sur {f['name']} ---\n{str(e)}"
|
||||
|
||||
file_analyses = []
|
||||
for f in extracted_files:
|
||||
analysis = await process_single_file(f)
|
||||
file_analyses.append(analysis)
|
||||
# SÉQUENÇAGE : Pause de 1.5 seconde entre les fichiers pour éviter l'erreur 429
|
||||
await asyncio.sleep(1.5)
|
||||
|
||||
files_context = "".join(file_analyses)
|
||||
user_prompt = f"{original_user_text}\n\nVoici les données pré-traitées des fichiers joints :\n{files_context}"
|
||||
|
||||
# --- OPTIMISATION & TRADUCTION ---
|
||||
log_activity(f"Optimisation/Traduction de la requête via Prompteur...")
|
||||
prompt_system = "You are an expert prompt engineer. Translate and optimize the user request and any file context into clear, precise English tailored for AI execution. Keep all code blocks intact."
|
||||
optimized = await ask_agent(p_client, p_mod, [{"role": "system", "content": prompt_system}, {"role": "user", "content": user_prompt}], p_prov)
|
||||
|
||||
log_activity(f"Lancement de {len(workers)} travailleur(s) en parallèle...")
|
||||
if len(workers) == 1:
|
||||
w_mod = workers[0]
|
||||
client, provider = get_client_for_model(db, w_mod, settings)
|
||||
worker_response = await ask_agent(client, w_mod, formatted_history + [{"role": "user", "content": optimized}], provider)
|
||||
responses = [worker_response]
|
||||
log_activity(f"Travailleur 1 ({w_mod}) a terminé.")
|
||||
else:
|
||||
w_tasks = []
|
||||
for w in workers:
|
||||
w_client, w_prov = get_client_for_model(db, w, settings)
|
||||
w_tasks.append(ask_agent(w_client, w, formatted_history + [{"role": "user", "content": optimized}], w_prov))
|
||||
responses = await asyncio.gather(*w_tasks, return_exceptions=True)
|
||||
log_activity("Tous les travailleurs ont terminé leur analyse.")
|
||||
|
||||
c_mod = config.get("concatenator")
|
||||
if not c_mod: c_mod = "gemini-3.5-flash-lite"
|
||||
|
||||
c_client, c_prov = get_client_for_model(db, c_mod, settings)
|
||||
|
||||
log_activity(f"Synthèse et traduction finale via Concaténeur ({c_mod})...")
|
||||
concat_system = (
|
||||
"You are a master lead developer and technical synthesizer. "
|
||||
"Synthesize the provided expert responses into a single cohesive response. "
|
||||
@@ -285,9 +306,12 @@ async def run_orchestrator(db: Session, history: list, settings: SystemSettings,
|
||||
|
||||
synth = f"User Request: {original_user_text}\n\n" + "\n".join([f"--- EXPERT {i+1} ---\n{str(r)}" for i, r in enumerate(responses)])
|
||||
final_response = await ask_agent(c_client, c_mod, [{"role": "system", "content": concat_system}, {"role": "user", "content": synth}], c_prov)
|
||||
|
||||
log_activity("Orchestration terminée avec succès.")
|
||||
|
||||
except Exception as e:
|
||||
final_response = f"L'IA a rencontré une erreur critique: {str(e)}"
|
||||
log_activity(f"ERREUR CRITIQUE: {str(e)}")
|
||||
|
||||
await sync_finances(db, settings)
|
||||
return final_response
|
||||
Reference in new issue
Block a user