Repository navigation
Expand file tree
/
Copy pathconfig.example.yml
More file actions
70 lines (61 loc) · 2.44 KB
/
Copy pathconfig.example.yml
File metadata and controls
70 lines (61 loc) · 2.44 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
# podwash configuration
# Copy to config.yml and customize
# Base URL where this service is accessible (used in RSS feed URLs)
base_url: "http://localhost:8080"
# Directory for SQLite database and audio files
data_dir: "./data"
# Feeds to monitor
feeds:
- name: "My Favorite Podcast"
url: "https://feeds.example.com/podcast.xml"
slug: "my-favorite-podcast"
poll_interval_minutes: 60
# Set false to publish placeholders without automatically queueing
# new episodes. They enter the worker queue only when requested.
auto_process: true
# - name: "Another Podcast"
# url: "https://feeds.example.com/another.xml"
# slug: "another-podcast"
# poll_interval_minutes: 120
# Optional: keep only source episodes whose title contains any of
# these case-insensitive substrings. Useful when a podcast host
# publishes multiple shows into one shared feed.
# title_includes:
# - "Midnight Boys"
# Processing settings
processing:
# Whisper model: tiny, base, small, medium, large-v3
whisper_model: "base"
# Whisper compute type: int8, float16, float32
whisper_compute_type: "int8"
# Minimum confidence to cut an ad segment (0.0-1.0)
confidence_threshold: 0.7
# Padding in seconds around ad boundaries
ad_boundary_padding: 0.5
# Maximum episodes to keep processed audio for (older get cleaned up)
retention_days: 30
# Classifier backend: "claude" (Anthropic SDK) or "litellm" (any
# OpenAI-compatible /chat/completions endpoint — a LiteLLM proxy,
# Ollama, vLLM)
classifier_backend: "claude"
# Claude API settings
claude:
model: "claude-sonnet-4-5-20250929"
max_tokens: 4096
# OpenAI-compatible backend (used when classifier_backend = "litellm").
# base_url is resolved on the worker host, not the server host.
litellm:
model: "dsf"
base_url: "http://localhost:4000/v1"
# A reasoning model sizes its answer to this budget rather than merely
# truncating at it, so a small value quietly costs ad segments while
# still returning valid JSON. Keep it generous — real output is ~350
# tokens, so this bounds quality, not spend.
max_tokens: 16000
# Optional bearer token for a proxy that requires one.
api_key: ""
# Self-healing alerts (set in .env on each host, NOT here):
# ALERT_TELEGRAM_BOT_TOKEN bot token
# ALERT_TELEGRAM_CHAT_ID target chat (e.g. -100...)
# ALERT_TELEGRAM_THREAD_ID optional thread id
# When unset, alerts still print to stderr but no Telegram message is sent.