saearch / src /config.py
xia
chore: CI tag latest + requirements mcp-proxy + NB10 retrain artefacts
a9be118
Raw History Blame Contribute Delete
8.16 kB
"""
Configuration centralisΓ©e du projet RAG Catastrophes Climatiques.
Hyperparamètres, spécialisation des LLM par agent, monitoring LLMOps.
"""
import os
from langchain_anthropic import ChatAnthropic
# ── ModΓ¨les Anthropic disponibles ─────────────────────────────────────────
MODEL_HAIKU = "claude-haiku-4-5-20251001"
MODEL_SONNET = "claude-sonnet-4-20250514"
MODEL_OPUS = "claude-opus-4-20250514"
# ── Tarifs Anthropic ($ / 1M tokens) pour estimation de coΓ»t ──────────────
MODEL_PRICING = {
MODEL_HAIKU: {"input": 0.25, "output": 1.25},
MODEL_SONNET: {"input": 3.00, "output": 15.00},
MODEL_OPUS: {"input": 15.00, "output": 75.00},
}
# ── Configuration par type d'agent ────────────────────────────────────────
AGENT_CONFIGS = {
"orchestrator": {
"model": MODEL_SONNET,
"temperature": 0,
"max_tokens": 2048,
"description": "Orchestre les outils et génère la réponse finale",
},
"rag": {
"model": MODEL_SONNET,
"temperature": 0.2,
"max_tokens": 2048,
"description": "RΓ©pond en citant les sources du corpus climatique",
},
"meteo": {
"model": MODEL_HAIKU,
"temperature": 0,
"max_tokens": 1024,
"description": "Interprète les données météo (actuel/historique/prévisions)",
},
"web": {
"model": MODEL_HAIKU,
"temperature": 0.2,
"max_tokens": 1024,
"description": "SynthΓ©tise les actualitΓ©s issues de la recherche web",
},
"analyst": {
"model": MODEL_OPUS,
"temperature": 0.5,
"max_tokens": 4096,
"description": "Croise les donnΓ©es pour produire l'analyse de risque",
},
"chat": {
"model": MODEL_HAIKU,
"temperature": 0.7,
"max_tokens": 1024,
"description": "Conversation simple, salutations, hors-sujet",
},
}
# ── ModΓ¨le de fallback ────────────────────────────────────────────────────
FALLBACK_MODEL = MODEL_HAIKU
# ── ParamΓ¨tres du chunking ────────────────────────────────────────────────
CHUNK_SIZE = 1500
CHUNK_OVERLAP = 150
# ── ParamΓ¨tres des embeddings ─────────────────────────────────────────────
EMBEDDING_MODEL = "paraphrase-multilingual-MiniLM-L12-v2"
FAISS_STORE_PATH = "faiss_store"
# ── ParamΓ¨tres du retriever ───────────────────────────────────────────────
RETRIEVER_K = 12
RETRIEVER_FETCH_K = 40
RETRIEVER_SEARCH_TYPE = "mmr"
RETRIEVER_LAMBDA = 0.7 # privilegie pertinence (1.0) vs diversite (0.0)
# ── Recherche hybride BM25 + Dense ────────────────────────────────────────
BM25_WEIGHT = 0.5
DENSE_WEIGHT = 0.5
# ── MΓ©moire conversationnelle ─────────────────────────────────────────────
MEMORY_WINDOW_SIZE = 20
# ── Monitoring LLMOps ─────────────────────────────────────────────────────
TOKEN_TRACKING = True
class TokenCounter:
"""Compteur de tokens par agent, par session, avec estimation de coΓ»t."""
def __init__(self):
self.total_input = 0
self.total_output = 0
self.calls_by_agent = {}
self.tokens_by_agent = {}
def log(self, agent_type: str, response):
usage = response.usage_metadata
input_tokens = usage["input_tokens"]
output_tokens = usage["output_tokens"]
self.total_input += input_tokens
self.total_output += output_tokens
self.calls_by_agent[agent_type] = self.calls_by_agent.get(agent_type, 0) + 1
if agent_type not in self.tokens_by_agent:
self.tokens_by_agent[agent_type] = {"input": 0, "output": 0}
self.tokens_by_agent[agent_type]["input"] += input_tokens
self.tokens_by_agent[agent_type]["output"] += output_tokens
def estimate_cost(self) -> float:
"""Estime le coΓ»t total en $ basΓ© sur les tarifs Anthropic."""
total_cost = 0.0
for agent_type, tokens in self.tokens_by_agent.items():
model = AGENT_CONFIGS.get(agent_type, {}).get("model", MODEL_HAIKU)
pricing = MODEL_PRICING.get(model, MODEL_PRICING[MODEL_HAIKU])
cost_input = (tokens["input"] / 1_000_000) * pricing["input"]
cost_output = (tokens["output"] / 1_000_000) * pricing["output"]
total_cost += cost_input + cost_output
return total_cost
def summary(self) -> dict:
return {
"total_input_tokens": self.total_input,
"total_output_tokens": self.total_output,
"total_tokens": self.total_input + self.total_output,
"calls_by_agent": self.calls_by_agent,
"tokens_by_agent": self.tokens_by_agent,
"estimated_cost_usd": self.estimate_cost(),
}
def reset(self):
self.total_input = 0
self.total_output = 0
self.calls_by_agent = {}
self.tokens_by_agent = {}
# ── API mΓ©tΓ©o OpenMeteo ───────────────────────────────────────────────────
OPENMETEO_BASE_URL = "https://api.open-meteo.com/v1"
OPENMETEO_GEOCODING_URL = "https://geocoding-api.open-meteo.com/v1/search"
METEO_FORECAST_DAYS = 7
# ── Chainlit ──────────────────────────────────────────────────────────────
CHAINLIT_HOST = "0.0.0.0"
CHAINLIT_PORT = 8000
# ── Construction du LLM ───────────────────────────────────────────────────
def get_llm(agent_type: str) -> ChatAnthropic:
"""Retourne un LLM configurΓ© pour un type d'agent donnΓ©."""
if agent_type not in AGENT_CONFIGS:
raise ValueError(
f"Type d'agent inconnu : '{agent_type}'. "
f"Disponibles : {list(AGENT_CONFIGS.keys())}"
)
if not os.getenv("ANTHROPIC_API_KEY"):
raise EnvironmentError(
"ANTHROPIC_API_KEY manquante. Ajoutez-la dans votre fichier .env"
)
config = AGENT_CONFIGS[agent_type]
return ChatAnthropic(
model=config["model"],
temperature=config["temperature"],
max_tokens=config["max_tokens"],
)
def get_fallback_llm() -> ChatAnthropic:
"""LLM de secours Anthropic (Haiku) en cas d'échec du modèle principal."""
return ChatAnthropic(
model=FALLBACK_MODEL,
temperature=0.2,
max_tokens=1024,
)
def get_ollama_fallback():
"""
LLM de secours local via Ollama (open source, gratuit, hors ligne).
Nécessite Ollama installé + modèle téléchargé : ollama pull mistral
UtilisΓ© quand l'API Anthropic est indisponible.
"""
try:
import langchain_community.llms as community_llms
ollama_cls = getattr(community_llms, "Ollama")
return ollama_cls(
model="mistral",
temperature=0.2,
num_predict=1024,
)
except (ImportError, AttributeError) as exc:
raise EnvironmentError(
"Ollama non disponible. Installez langchain-community et Ollama : "
"pip install langchain-community && curl -fsSL https://ollama.com/install.sh | sh && ollama pull mistral"
) from exc