config.py

import os

# --- API Keys & Secrets ---
# GOOGLE_API_KEY    = os.environ.get("GOOGLE_API_KEY", "")
PINECONE_API_KEY  = os.environ.get("PINECONE_API_KEY", "")
GOOGLE_CLIENT_ID  = os.environ.get("GOOGLE_CLIENT_ID", "")

# Optional: shared fallback key (users bring their own via BYOK)
OPENROUTER_API_KEY = os.environ.get("OPENROUTER_API_KEY", "")

# Optional: shared Tavily web-search fallback key (users bring their own via BYOK)
TAVILY_API_KEY = os.environ.get("TAVILY_API_KEY", "")

# --- LangSmith (tracing & analytics) ---
os.environ["LANGCHAIN_TRACING_V2"] = os.environ.get("LANGSMITH_TRACING", "true")
os.environ["LANGCHAIN_API_KEY"] = os.environ.get("LANGSMITH_API_KEY", "")
os.environ["LANGCHAIN_PROJECT"] = os.environ.get("LANGCHAIN_PROJECT", "stemcopilot-canvas")

# --- Pinecone ---
PINECONE_INDEX = "stem-embed"

# --- Embedding ---
EMBED_MODEL_NAME  = "BAAI/bge-large-en-v1.5"
BGE_QUERY_PREFIX  = "Represent this sentence: "

# --- Database ---
DB_PATH = "/data/stemgraph.db" if os.path.isdir("/data") else "stemgraph.db"

# Public URL of the deployed app — used as the OpenRouter HTTP-Referer header.
APP_URL = os.environ.get("APP_URL", "https://moai633-stem-copilot.hf.space")

# --- Defaults ---
DEFAULT_PERSONA = "nerd"
RETRIEVAL_TOP_K = 3
LLM_TIMEOUT = 180  # seconds — free models are slow, needs generous timeout

# Number of most-recent messages sent to the LLM per request (full history is
# still stored for the /history view). Caps input-token growth on long chats.
CONTEXT_WINDOW = int(os.environ.get("CONTEXT_WINDOW", "10"))

# Minimum Pinecone relevance score for a chunk to be injected as context.
RETRIEVAL_MIN_SCORE = float(os.environ.get("RETRIEVAL_MIN_SCORE", "0.45"))

# --- Logging ---
import logging
logging.basicConfig(
    level=os.environ.get("LOG_LEVEL", "INFO").upper(),
    format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
)