backend/core/config.py

"""
core/config.py — Centralized settings via pydantic-settings.
All values read from environment variables (set in HF Spaces secrets).
"""

from enum import Enum
from functools import lru_cache

from pydantic import Field, computed_field
from pydantic_settings import BaseSettings, SettingsConfigDict


class HighlightColor(str, Enum):
    GREEN = "green"    # Fact-checked, widely corroborated
    YELLOW = "yellow"  # Breaking / unverified / weak signal
    RED = "red"        # Debunked, active community note
    PURPLE = "purple"  # LLM hallucination detected


class Platform(str, Enum):
    TWITTER = "twitter"
    INSTAGRAM = "instagram"
    YOUTUBE = "youtube"
    CHATGPT = "chatgpt"
    CLAUDE = "claude"
    GEMINI = "gemini"
    NEWS = "news"
    UNKNOWN = "unknown"


class Settings(BaseSettings):
    model_config = SettingsConfigDict(env_file=".env", extra="ignore")

    # LLM API keys
    groq_api_key: str = Field(default="", alias="GROQ_API_KEY")
    x_bearer_token: str = Field(default="", alias="X_BEARER_TOKEN")

    # Infrastructure
    qdrant_host: str = Field(default="localhost", alias="QDRANT_HOST")
    qdrant_port: int = Field(default=6333, alias="QDRANT_PORT")
    memgraph_host: str = Field(default="localhost", alias="MEMGRAPH_HOST")
    memgraph_port: int = Field(default=7687, alias="MEMGRAPH_PORT")
    memgraph_password: str = Field(default="memgraph123", alias="MEMGRAPH_PASSWORD")
    redpanda_brokers: str = Field(default="localhost:9092", alias="REDPANDA_BROKERS")
    redis_url: str = Field(default="redis://localhost:6379", alias="REDIS_URL")

    # App
    port: int = Field(default=7860, alias="PORT")
    log_level: str = Field(default="INFO", alias="LOG_LEVEL")
    demo_mode: bool = Field(default=False, alias="DEMO_MODE")

    # Model identifiers for LiteLLM routing
    gatekeeper_model: str = "groq/llama3-8b-8192"
    misinformation_model: str = "groq/mixtral-8x7b-32768"
    hallucination_model: str = "groq/llama3-70b-8192"  # Free via Groq — replaces Claude Haiku

    # Gatekeeper latency SLO: p95 < 120ms
    gatekeeper_timeout_ms: int = 120

    # Cache TTLs (seconds)
    cache_ttl_green_red: int = 21_600   # 6 hours
    cache_ttl_yellow: int = 900          # 15 minutes
    # Purple: no cache — hallucination checks are context-specific

    # RAG retrieval
    qdrant_collection: str = "claims"
    qdrant_ef: int = 128        # HNSW ef parameter — higher = more accurate, slower
    qdrant_top_k: int = 8       # nearest neighbors to retrieve
    evidence_window_hours: int = 72  # only retrieve evidence newer than 72h

    # Minimum text length for analysis (words)
    min_word_count: int = 12

    @computed_field
    @property
    def has_groq(self) -> bool:
        return bool(self.groq_api_key)

    @computed_field
    @property
    def has_hallucination_llm(self) -> bool:
        # Hallucination agent uses Groq llama3-70b (free) — same key as gatekeeper
        return bool(self.groq_api_key)

    @computed_field
    @property
    def has_x_api(self) -> bool:
        return bool(self.x_bearer_token)

    @computed_field
    @property
    def broker_list(self) -> list[str]:
        return self.redpanda_brokers.split(",")


@lru_cache
def get_settings() -> Settings:
    return Settings()