# Copy this file to .env and fill in your real values.
# .env is never committed — only this .example file is.
# Every command needed to provision the cloud side is in commands/commands.md.

PROJECT_ID=your-gcp-project-id
LOCATION=us-central1     # Vertex AI region (read by the app)
REGION=us-central1       # Cloud Run deploy region (used by commands/commands.md / gcloud only)

# Cloud Storage — where the raw HR policy documents live
GCS_BUCKET_NAME=your-bucket-name

# Jina AI — embeddings + reranker
JINA_API_KEY=your-jina-api-key-here

# Models — optional overrides (defaults in hr_assistant/config.py).
# gemini-2.5-flash retires ~2026-10-20 — set LLM_MODEL_NAME to a Gemini 3.x
# Flash model before then (no code change — llm.py prefixes it "vertex_ai/").
# FALLBACK_MODEL_NAME is the Groq model llm.py's LiteLLM router falls back to
# if the Vertex call errors (prefixed "groq/"). Changing EMBEDDING_MODEL_NAME
# also needs `python ingest.py --force` (vector dimension changes).
# LLM_MODEL_NAME=gemini-2.5-flash
# FALLBACK_MODEL_NAME=openai/gpt-oss-20b
# EMBEDDING_MODEL_NAME=jina-embeddings-v2-base-en
# RERANKER_MODEL_NAME=jina-reranker-v2-base-multilingual

# Qdrant Cloud — the vector store
QDRANT_URL=https://your-cluster-url.qdrant.io
QDRANT_API_KEY=your-qdrant-api-key-here
QDRANT_COLLECTION_NAME=hr_policies

# --- Guardrails / cache (the secure pipeline app.py + main.py run) ---

# Separate Qdrant collection holding HR docs + the noise corpus together
# (used by demo_reliability.py / redteam_test.py)
QDRANT_NOISY_COLLECTION_NAME=hr_policies_noisy_demo

# Input/output safety guardrail — runs on EVERY request now. "model_armor"
# (default, needs the template provisioned — see commands/commands.md Phase 5),
# "gemini_lite" (one Gemini call, no extra GCP setup — good for local dev),
# or "none" (disabled).
GUARDRAIL_PROVIDER=model_armor
MODEL_ARMOR_LOCATION=us
MODEL_ARMOR_TEMPLATE_ID=hr-assistant-guardrail

# Optional tuning — all have sensible defaults; leave unset unless tuning.
#   GUARDRAIL_FAIL_OPEN_INPUT=false    # on a provider ERROR: input fails closed
#   GUARDRAIL_FAIL_OPEN_OUTPUT=true    #                       output fails open
#   GUARDRAIL_HISTORY_TURNS=6          # recent USER turns screened with each question (0 = off)
#   SEMANTIC_CACHE_MAX_ENTRIES=500     # oldest evicted past this
#   SEMANTIC_CACHE_TTL_SECONDS=3600    # cache entry lifetime (0 = never expire)

# LangSmith — tracing for the app + datasets/experiments for evaluate.py.
# Free API key at smith.langchain.com. Set LANGSMITH_TRACING=true to send
# traces; the app shows a green/grey status line in the sidebar.
LANGSMITH_TRACING=false
LANGSMITH_ENDPOINT=https://api.smith.langchain.com
LANGSMITH_API_KEY=your-langsmith-api-key-here
LANGSMITH_PROJECT=hr-policy-assistant

# Groq — used two places: the evaluate.py judge (gpt-oss-120b, a different
# family from the app's Gemini) and the app's fallback model
# (FALLBACK_MODEL_NAME, if the Vertex Gemini call errors). Free key at
# console.groq.com. The app runs without it — the fallback just can't fire.
GROQ_API_KEY=your-groq-api-key-here
# GROQ_BASE_URL=https://api.groq.com/openai/v1
# JUDGE_MODEL_NAME=openai/gpt-oss-120b