.env.example

BRAVE_API_KEY=
GOOGLE_SAFE_BROWSING_KEY=

# ---------------------------------------------------------------------------
# LLM judge — the core vectorizer layer.
#
# DEFAULT: OpenRouter + Groq-routed Llama 3.3 70B
#   - Measured ~3s/call average, 11-12/12 accuracy on the validation set
#   - Cost ~$0.22 per 1000 calls
#   - Get a key: https://openrouter.ai/keys
#
# ALTERNATIVE: local LLM (LM Studio / Ollama / llama-server / vLLM)
#   - Set LLM_ENDPOINT=http://localhost:1234/v1 (or your local server)
#   - Leave LLM_API_KEY blank
#   - LLM_PROVIDER_ORDER is ignored by local backends
# ---------------------------------------------------------------------------

LLM_JUDGE_ENABLED=true
LLM_ENDPOINT=https://openrouter.ai/api/v1
LLM_MODEL=meta-llama/llama-3.3-70b-instruct
LLM_API_KEY=
# OpenRouter provider routing: comma-separated preferred providers.
# Leave empty to let OpenRouter auto-route with fallback — this is more
# resilient under concurrent load. "groq" gives the fastest per-call
# latency when the provider isn't throttled, but concurrent bursts can
# stall on single-provider routing. Set only if you're tuning for
# low-QPS, low-latency workloads.
# LLM_PROVIDER_ORDER=groq
LLM_TIMEOUT_MS=45000
# LLM_CONCURRENCY: parallel in-flight requests to the LLM endpoint.
# OpenRouter / hosted: 4-8 works well. Local LM Studio: 2 is the RTX 5060 Ti 16GB optimum.
LLM_CONCURRENCY=4
# Thinking mode: prepends a reasoning preamble and raises the token budget.
# For modern hosted models this is usually unnecessary and adds latency.
# Enable only when using a local model that needs explicit reasoning traces.
LLM_THINKING=false

# Server-wide localization defaults — overridden by tool input
# SIFT_DEFAULT_COUNTRY=US      # 2-char uppercase ISO
# SIFT_DEFAULT_SEARCH_LANG=en  # 2-3 char lowercase, normalized per backend

# Observation logging — append-only local file.
# Set to "off" to fully disable.
# SIFT_OBSERVATIONS=on

# ---------------------------------------------------------------------------
# Optional: mirror observations to a remote HTTP store (learning-loop
# substrate for drift / accuracy review).
#
# Opt-in. Unset SIFT_OBSERVATION_SYNC_URL to disable (default).
# The local JSONL file is unaffected — this is purely additive.
#
# Each MCP call appends to a daily file at:
#   {URL}/{PATH}YYYY-MM-DD.jsonl
# via GET → append → PUT (read-modify-write).
#
# Works with any HTTP PUT-capable endpoint. Example — Obsidian Local REST API:
#   SIFT_OBSERVATION_SYNC_URL=https://127.0.0.1:27124/vault
#   SIFT_OBSERVATION_SYNC_AUTH=Bearer <your-obsidian-api-token>
#   SIFT_OBSERVATION_SYNC_PATH=observations/sift/
#   SIFT_OBSERVATION_SYNC_INSECURE_TLS=true   # Obsidian uses a self-signed cert
#
# Also compatible with: WebDAV servers, S3/GCS signed URLs (short-lived),
# Gitea/Gitlab raw API (with auth), custom log collectors.
# ---------------------------------------------------------------------------
# SIFT_OBSERVATION_SYNC_URL=
# SIFT_OBSERVATION_SYNC_AUTH=
# SIFT_OBSERVATION_SYNC_PATH=observations/sift/
# SIFT_OBSERVATION_SYNC_INSECURE_TLS=false