sidecar/config.py

"""
sidecar/config.py — Centralised configuration for the GenAI Shield Sidecar.

All runtime behaviour is controlled via environment variables.
Defaults are safe for local development.
"""

import os

# ── Sidecar server ────────────────────────────────────────────────────────────
SIDECAR_HOST     = os.getenv("SIDECAR_HOST", "0.0.0.0")
SIDECAR_PORT     = int(os.getenv("SIDECAR_PORT", "5050"))

# ── LLM backend ───────────────────────────────────────────────────────────────
LLM_BACKEND      = os.getenv("LLM_BACKEND", "gemini")   # "gemini" | "openai"
GEMINI_API_KEY   = os.getenv("GEMINI_API_KEY", "")
GEMINI_MODEL     = os.getenv("GEMINI_MODEL", "gemini-3.1-flash-lite")

OPENAI_API_KEY   = os.getenv("OPENAI_API_KEY", "")
OPENAI_BASE_URL  = os.getenv("OPENAI_BASE_URL", "https://api.openai.com/v1")
OPENAI_MODEL     = os.getenv("OPENAI_MODEL", "gpt-4o-mini")

# ── System prompt ─────────────────────────────────────────────────────────────
SYSTEM_PROMPT = os.getenv(
    "GENAI_SYSTEM_PROMPT",
    "You are a helpful AI assistant. Be concise, accurate, and professional.",
)

# ── Shield thresholds ─────────────────────────────────────────────────────────
GUARD_BLOCK_THRESHOLD   = float(os.getenv("GUARD_BLOCK_THRESHOLD",   "0.85"))
GUARD_FLAG_THRESHOLD    = float(os.getenv("GUARD_FLAG_THRESHOLD",    "0.50"))
MONITOR_BLOCK_THRESHOLD = int(os.getenv("MONITOR_BLOCK_THRESHOLD",  "40"))

# ── Gate behaviour ────────────────────────────────────────────────────────────
# Max ms to wait for guard verdict before letting buffered tokens through anyway
# (safety valve — guard should always be <300ms in practice)
GATE_GUARD_TIMEOUT_SEC  = float(os.getenv("GATE_GUARD_TIMEOUT_SEC", "3.0"))

# ── Post-monitor thread pool ──────────────────────────────────────────────────
MONITOR_WORKERS         = int(os.getenv("MONITOR_WORKERS", "4"))

# ── Sentence splitter ─────────────────────────────────────────────────────────
# Minimum characters before a sentence boundary is declared
SENTENCE_MIN_CHARS      = int(os.getenv("SENTENCE_MIN_CHARS", "20"))

# ── Model path ────────────────────────────────────────────────────────────────
PROMPT_GUARD_MODEL_DIR  = os.getenv(
    "PROMPT_GUARD_MODEL_DIR",
    "models/Llama-Prompt-Guard-2-86M",
)

# ── Logging ───────────────────────────────────────────────────────────────────
LOG_LEVEL = os.getenv("LOG_LEVEL", "INFO")