promptforge/workshop.example.toml at master · cppalliance/promptforge · GitHub

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
71
# PromptForge Workshop example configuration.
#
# The promptforge-ws desktop shell generates ~/.promptforge/workshop.toml
# from a built-in template on first run, so no setup is needed there. This
# file is the commented reference: copy it to workshop.toml and edit when
# you want a configuration the template does not produce - beside the
# shell's executable or in the current directory to shadow the profile
# copy, or in the current directory for the promptforge-ws-server binary,
# which reads workshop.toml from there (or workbench.toml if that is absent).
#
# String values support ${VAR} environment interpolation; $$ is a literal
# $, and an unset variable interpolates to the empty string.

[gateway]
# Base URL of the PromptForge gateway.
# Default when empty: "http://127.0.0.1:8081"
base_url = "http://127.0.0.1:8081"
# Bearer key for the gateway API. Empty sends no Authorization header.
api_key = "${PROMPTFORGE_GATEWAY_API_KEY}"

[tape]
# Path of the JSONL session tape; one event per chat exchange.
# Default: "tape.jsonl"
path = "tape.jsonl"

[server]
# Address the workshop server binds to.
# Default: "127.0.0.1:7910"
bind = "127.0.0.1:7910"
# When true, the server binary opens the system browser at its address once
# serving. The desktop shell ignores this flag; it exists for the
# browser-tab frame.
# Default: false
open_browser = false

[voice]
# Path to the GGML whisper model for streaming interim transcription.
# Empty disables transcription until a source provides the model; the
# /voice endpoint then captures and counts PCM but emits empty
# transcripts. A missing file with interim_source set is downloaded
# through the gateway cache once the gateway connects; a missing file
# with no source degrades to voice disabled (with a status-bar note),
# never a startup failure. Models come from
# https://huggingface.co/ggerganov/whisper.cpp (for example
# ggml-large-v3-turbo.bin for interim, ggml-large-v3.bin for final).
interim_model = "models/ggml-large-v3-turbo.bin"
# Path to the whisper model for the pipelined final pass over a take.
# Empty falls back to one last interim-model window; a missing file with
# no final_source drops just the final pass.
final_model = "models/ggml-large-v3.bin"
# URL the interim model is downloaded from through the gateway cache when
# no local file exists.
# Default: "" (no known source)
interim_source = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo.bin"
# URL the final-pass model is downloaded from through the gateway cache
# when no local file exists.
# Default: "" (no known source)
final_source = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3.bin"
# Seconds of trailing audio each interim pass transcribes.
# Default: 5
window_seconds = 5
# Milliseconds between interim passes while a take is recording.
# Default: 800
interval_ms = 800
# Domain terms whisper is biased toward (for example project jargon the
# model would otherwise mishear). The terms are formatted into a glossary
# conditioning prompt on both the interim and final-pass workers; the
# glossary is truncated to fit whisper's 224-token prompt budget, with a
# warning logged when terms are dropped.
# Default: [] (no biasing)
# vocabulary = ["MCP", "GGUF", "Lua"]