logan/server-26
1
0
Fork 0
Code Issues 84 Pull Requests Actions Packages Projects Releases Wiki Activity

Docs: document the STT provider, extraction and alert-gate env vars (server-26#297, server-26#298, server-26#300)

Browse Source 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Logan Cusano
2026-10-27 20:29:10 -04:00
co-authored by Claude Opus 5.5
parent ff0a89e9f1
commit 90a3c27a24
1 changed files with 42 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
drb-c2-core/.env.example
+42 -1
View File
@@ -36,8 +36,49 @@ NWS_USER_AGENT=DRB (you@example.com)
# Comma list of NWS state/marine area codes served
NWS_AREA=NY
# OpenAI — for transcription (Whisper), intelligence extraction, embeddings, and summaries
# OpenAI — transcription (whisper-1, and the STT fallback), extraction, embeddings
OPENAI_API_KEY=
# OpenAI transcription model: whisper-1 | gpt-4o-mini-transcribe | gpt-4o-transcribe
STT_MODEL=whisper-1
# STT provider (server-26#297). Any OpenAI-compatible /audio/transcriptions API.
# auto (default): groq if GROQ_API_KEY is set, compat if STT_BASE_URL is set,
# otherwise openai — so leaving these blank keeps today's OpenAI behaviour.
STT_PROVIDER=auto
# Groq: whisper-large-v3-turbo, $0.04/audio-hour (min 10s billed per request).
GROQ_API_KEY=
# Model on the non-OpenAI provider.
STT_COMPAT_MODEL=whisper-large-v3-turbo
# Override the endpoint/key (blank = provider default: https://api.groq.com/openai/v1
# and GROQ_API_KEY). Set both for a self-hosted OpenAI-compatible server.
STT_BASE_URL=
STT_API_KEY=
# Retry on OpenAI STT_MODEL when the provider above fails (needs OPENAI_API_KEY).
STT_FALLBACK_TO_OPENAI=true
# Area places / local names / vocabulary in the STT prompt (non-OpenAI provider only).
STT_VOCAB_PROMPT_ENABLED=true
# Estimated prompt token budget (Whisper keeps only the last 224).
STT_PROMPT_MAX_TOKENS=200
# Scene extraction / embeddings models (server-26#298)
EXTRACTION_MODEL=gpt-4o-mini
EMBEDDING_MODEL=text-embedding-3-small
# Skip correction + extraction for radio housekeeping (roll call, bare 10-4s).
# Default false = shadow mode (verdict recorded, nothing skipped). Turn on only
# after a shadow window shows no real events flagged (server-26#127).
CHATTER_GATE_ENABLED=false
# In-process geocode cache (Maps terms allow caching lat/lng <= 30 days)
GEOCODE_CACHE_TTL_SECONDS=86400
GEOCODE_CACHE_NEGATIVE_TTL_SECONDS=3600
GEOCODE_CACHE_MAX_ENTRIES=5000
# Push-alert confidence gate (server-26#300). Below the floor -> held (admin-only
# GET /alerts/held). High-severity matches also need N calls on the incident or
# the high confidence bar.
ALERT_MIN_CONFIDENCE=0.5
ALERT_HIGH_SEVERITY_MIN_CONFIDENCE=0.85
ALERT_CORROBORATION_MIN_CALLS=2
ALERT_STT_REFERENCE=0.3
SUMMARY_INTERVAL_MINUTES=15
CORRELATION_WINDOW_HOURS=4
EMBEDDING_SIMILARITY_THRESHOLD=0.82