src/config.py

"""Application configuration.

Loads settings and secrets from environment variables (and the local .env file)
using pydantic-settings. Everything the rest of the app needs — API keys, model
names, the SQLite path, generation tunables — lives here, so no other module
ever has to touch os.environ directly.

Usage:
    from src.config import settings
    settings.anthropic_api_key
"""

from __future__ import annotations

import os

from pydantic_settings import BaseSettings, SettingsConfigDict

# pydantic-settings ranks real environment variables ABOVE the .env file. Some
# shells/CI (including the dev environment this was built in) export an *empty*
# ANTHROPIC_API_KEY, which would silently shadow the real value in .env. Drop
# any of our secrets that are present-but-empty so .env can fill them in. This
# is safe on Hugging Face Spaces, where secrets arrive as non-empty env vars and
# therefore still take priority.
_SECRET_ENV_VARS = (
    "ANTHROPIC_API_KEY",
    "HF_TOKEN",
    "LANGFUSE_PUBLIC_KEY",
    "LANGFUSE_SECRET_KEY",
    "LANGFUSE_HOST",
    "TAVILY_API_KEY",
)
for _key in _SECRET_ENV_VARS:
    if os.environ.get(_key, None) == "":
        del os.environ[_key]


class Settings(BaseSettings):
    """Typed view over the .env file / environment variables.

    Secrets default to empty strings so that *importing* this module never
    fails just because a key is missing. Each assistant/tool checks that the
    key it needs is actually present at call time and raises a clear error if
    not. This keeps the smoke tests and partial setups working.
    """

    model_config = SettingsConfigDict(
        env_file=".env",
        env_file_encoding="utf-8",
        extra="ignore",  # ignore unrelated env vars instead of erroring
    )

    # --- Secrets (loaded from .env) ---
    anthropic_api_key: str = ""
    hf_token: str = ""
    langfuse_public_key: str = ""
    langfuse_secret_key: str = ""
    langfuse_host: str = "https://cloud.langfuse.com"
    tavily_api_key: str = ""

    # --- Model identifiers ---
    # Frontier assistant (and the eval judge) use Claude Sonnet 4.5.
    frontier_model: str = "claude-sonnet-4-5"
    # Output-moderation guardrail uses the cheaper/faster Haiku.
    moderation_model: str = "claude-haiku-4-5"
    # Open-source assistant.
    oss_model: str = "Qwen/Qwen2.5-1.5B-Instruct"

    # --- Generation tunables (shared by both assistants for fair comparison) ---
    max_tokens: int = 512
    temperature: float = 0.7

    # --- Storage ---
    # SQLite file backing the per-session chat memory (added in Phase 3).
    sqlite_path: str = "./data/sessions.db"


# Single shared instance imported across the app.
settings = Settings()