# ╔══════════════════════════════════════════════════════════╗

# ║ ProjectGabriel Configuration ║

# ╚══════════════════════════════════════════════════════════╝

# Copy this file to config.yml and fill in your values.

# Display name used in the WebUI, console logs, OBS overlays, and chatbox.

app_name: "Gabriel"

# Which brain to run. Two options:

# gemini_live - cloud Gemini Live websocket (the original setup, native voice)

# local - LM Studio for the LLM + Moonshine v2 for STT + your chosen

# external TTS provider (qwen3 / hoppou / chirp3hd / tiktok / plugin)

# Local mode is fully offline aside from a few short flash-lite sub-agents

# (memory recall, conversation summary). Local mode REQUIRES one of the

# external TTS providers under tts:* to be enabled, since LM Studio has no

# native voice.

backend: "gemini_live"

# Privacy controls. Off by default. See README "Privacy" section.

privacy:

# When true, every Gemini Live session writes a JSON transcript of the

# whole conversation (user transcripts, assistant transcripts, tool

# calls, tool responses) to data/conversations/<timestamp>.json. Off by

# default. Turn on if you want a personal log to grep / replay later,

# leave off if you do not want any of it on disk.

save_conversations: false

# Logging. Controls how chatty the terminal is.

logging:

# DEBUG / INFO / WARNING / ERROR. Use DEBUG to see everything including

# third party libs (httpx, discord internals, etc), INFO is the everyday

# level, WARNING hides per-event status spam.

level: "INFO"

gemini:

# Primary API key (required) - get from https://aistudio.google.com/apikey

api_key: "YOUR_GEMINI_API_KEY_HERE"

# Backup keys for automatic rotation when rate limited (optional)

# The system will cycle through these when the primary key hits quota limits

backup_keys: []

# - "BACKUP_KEY_1"

# - "BACKUP_KEY_2"

# - "BACKUP_KEY_3"

# Gemini Live model - must support native audio

# Options: "gemini-3.1-flash-live-preview" (latest, lower latency)

# "gemini-2.5-flash-native-audio-preview-09-2025" (legacy)

# "gemini-2.5-flash-native-audio-preview-12-2025" (legacy)

model: "gemini-2.5-flash-native-audio-preview-09-2025"

# System prompt -- select a named prompt from config/prompts/prompts.yml

prompt: "default"

# ── Voice Configuration ──

# Prebuilt voice name. Available voices:

# Puck, Charon, Kore, Fenrir, Aoede, Leda, Orus, Zephyr

voice: "Puck"

# ── Voice Activity Detection (VAD) ──

# Controls how Gemini detects when you start/stop speaking.

#

# Two modes:

# 1. Automatic (disabled: false) - Server-side VAD handles everything.

# Simple, works well for most setups. Just configure sensitivity.

# 2. Manual (disabled: true) - Client-side energy-based VAD.

# Sends activityStart/activityEnd signals to Gemini manually.

# Recommended for production apps (per Google AI community patterns).

# Gives you full control over turn-taking and prevents echo issues.

# Also gates audio during tool calls and model speech to avoid

# 1007/1008 disconnects from stale audio.

vad:

# VAD mode: "auto" (Gemini server-side) or "silero" (local Silero VAD model)

# auto = Uses Gemini's built-in automatic activity detection. Simple, no extra deps.

# silero = Uses Silero VAD locally for speech detection with activityStart/activityEnd.

# More stable on 3.1 models, prevents stalls from noise. Requires torch.

# Also gates audio during tool calls and model speech to avoid

# 1007/1008 disconnects from stale audio leaking to the server.

mode: "auto"

# Sensitivity for detecting the START of speech (auto mode only)

# Options: START_SENSITIVITY_LOW, START_SENSITIVITY_HIGH

# HIGH = triggers faster but may pick up noise, LOW = more selective

start_of_speech_sensitivity: "START_SENSITIVITY_HIGH"

# Sensitivity for detecting the END of speech (auto mode only)

# Options: END_SENSITIVITY_LOW, END_SENSITIVITY_HIGH

# HIGH = ends speech detection quicker, LOW = waits longer

end_of_speech_sensitivity: "END_SENSITIVITY_HIGH"

# Milliseconds of audio to capture BEFORE speech is detected (context padding)

prefix_padding_ms: 100

# Milliseconds of silence required to consider speech complete

# Used in both auto and silero mode. 500ms is good for natural conversation.

silence_duration_ms: 200

# Speech probability threshold for Silero VAD (silero mode only)

# Silero outputs a 0.0-1.0 probability per audio chunk.

# Chunks above this threshold are considered speech.

# Default 0.5 is a good starting point. Lower = more sensitive, higher = stricter.

# silero_threshold: 0.5

# ── Generation Parameters ──

temperature: 1.0 # 0.0-2.0, controls randomness (default: 1.0, higher = more creative)

# ── Thinking Configuration ──

# Enables the model to "think" before responding (inner monologue).

# Requires a model that supports thinking (e.g. gemini-2.5-flash-native-audio-preview-12-2025).

# Dynamic thinking is enabled by default on supported models.

thinking:

# For 2.5 models: thinking token budget - guides how many tokens the model uses to think.

# Higher = more thorough but slower responses. Set to 0 to disable thinking.

# null = use model default (dynamic thinking)

budget: null

# For 3.1 models: thinking level controls thinking depth.

# Options: "minimal" (lowest latency), "low", "medium", "high"

# Default is "minimal" for 3.1 models. Ignored for 2.5 models.

# level: "low"

# Whether to include thought summaries in server responses.

# Useful for debugging what the model is reasoning about.

include_thoughts: false

# ── Advanced Live Features (require v1alpha API, 2.5 models only, auto-fallback if unsupported) ──

# These are NOT supported on 3.1 models and will be automatically skipped.

# enable_affective_dialog: true # Emotional/expressive responses

# proactivity: null # Model proactivity level (float), null = default

# ── Google Search Grounding ──

# Enable Google Search tool for the model to look up current information.

# Auto-detection: if not set, enabled for 2.5 models, disabled for 3.1 (quota issues).

# Set explicitly to override: true = always on, false = always off.

# google_search: true

# ── Context Window Compression ──

# Prevents session termination when context gets too large.

# Without compression: audio-only 15 min, audio+video 2 min.

# With compression: unlimited session length.

# NOTE: These settings are fully disabled when custom_compression is enabled.

# No built-in compression config is sent to Gemini in that case.

context_window_compression:

enabled: true

# Token count that triggers compression (null = 80% of model's 128k context window)

trigger_tokens: 50000

# Tokens to keep after compression (null = trigger_tokens / 2)

target_tokens: 25000

# ── Custom Context Compression ──

# When enabled, replaces Gemini's built-in sliding window compression with a

# smarter approach: when tokens near the threshold, a lightweight model summarizes

# the conversation, clears the session, and seeds the summary as initial context.

# This avoids repeated 1007 errors that Gemini's built-in compression can cause.

custom_compression:

enabled: false

# Token count that triggers the summarize-and-reconnect cycle

trigger_tokens: 100000

# Model used for summarization (lightweight, fast)

model: "gemini-3.1-flash-lite"

# ── Session Resumption ──

# Controls how aggressively session handles are cleared on errors and

# how many recent messages are replayed as context on fresh reconnects.

session:

# Number of consecutive errors (e.g. 1007) before clearing session handle and starting fresh.

# Lower = faster recovery but discards handle sooner. Default: 1.

error_threshold: 1

# Number of recent user/assistant messages to replay as context on fresh reconnect.

# These are sent to the new session so the model knows what was being discussed.

replay_messages: 10

# ══════════════════════════════════════════════════════════

# Local backend (LM Studio + Moonshine + external TTS)

# ══════════════════════════════════════════════════════════

# Only used when backend: local at the top of this file. Otherwise ignored.

# Requirements:

# 1. LM Studio running with a chat model loaded and server enabled

# (Developer tab -> Start Server). default endpoint is the one below.

# 2. One of the external TTS providers below (tts.qwen3 / hoppou /

# chirp3_hd / tiktok or a plugin) enabled. local mode refuses to

# start without a TTS provider since the LLM has no voice of its own.

# 3. Moonshine Voice installed (pip install moonshine-voice). First

# run downloads the chosen streaming arch into the user cache.

local:

llm:

# OpenAI compatible endpoint. LM Studio defaults to localhost:1234.

base_url: "http://localhost:1234/v1"

# Model identifier as shown in LM Studio. The default placeholder works

# because LM Studio routes to whatever is loaded, but be explicit when

# you can.

model: "local-model"

api_key: "lm-studio"

temperature: 0.8

top_p: 0.95

max_tokens: 1024

# Rolling chat history kept in context across turns. Higher = more

# memory but more tokens per request.

history_messages: 30

# Per-request timeout in seconds.

request_timeout: 120

# Safety cap on tool call -> response -> tool call -> response loops in

# a single user turn. If the model wants more iterations than this we

# stop and let it speak.

max_tool_iterations: 6

vision:

# When true, captures the screen and attaches an image to every user

# turn. Requires a multimodal model loaded in LM Studio (eg qwen2.5-vl).

enabled: false

# Longest edge of the resized image. Smaller = faster, fewer tokens.

max_size: 768

# JPEG quality (1-100) for the encoded image.

quality: 70

stt:

# Moonshine streaming model arch. Picks weights via moonshine_voice.

# Options:

# tiny_streaming 34M params, ~69ms TTFT, lowest accuracy

# small_streaming 123M params, ~165ms TTFT, recommended default

# medium_streaming 245M params, ~269ms TTFT, best accuracy

model: "small_streaming"

# Language tag passed to moonshine_voice.download_model. Only English

# has full streaming weights; other languages fall back to base v2.

language: "en"

# Drop any speech segment shorter than this. Filters out keyboard

# clicks, coughs, brief mouth noises.

min_speech_ms: 400

# Hard cap on a single utterance so a noisy room can't grow the buffer

# forever. The utterance is force-finalised at this length.

max_utterance_ms: 30000

# Keep this much audio from BEFORE VAD triggered so we never clip the

# first phoneme. 200-400ms is the sweet spot.

pre_roll_ms: 300

# Silero VAD speech-probability threshold (0.0-1.0). Lower = more

# sensitive (picks up quieter speech, more false triggers). Higher =

# stricter (might miss soft talkers). Defaults to gemini.vad.silero_threshold

# if unset, which itself defaults to 0.5. Try 0.3 if Gabriel keeps

# missing your voice, 0.6+ if the mic triggers on background noise.

vad_threshold: 0.4

# How many ms of silence before we close the utterance and send it to

# transcription. Lower = snappier turn-taking but cuts off mid-sentence

# pauses. Higher = more patient. Falls back to gemini.vad.silence_duration_ms.

silence_ms: 600

# ══════════════════════════════════════════════════════════

# TTS Provider

# ══════════════════════════════════════════════════════════

# Choose which voice to use for AI speech output.

# "gemini" = use Gemini Live's native audio (default)

# "qwen3" = discard Gemini audio, use Qwen3 TTS server instead

# "hoppou" = discard Gemini audio, use Hoppou AI cloud TTS

# "chirp3_hd" = discard Gemini audio, use Google Cloud Chirp 3: HD streaming TTS

# "tiktok" = discard Gemini audio, use TikTok's TTS API (free, no API key needed)

tts:

provider: "gemini"

# Providers the AI can switch to at runtime via the switchTTSProvider tool.

# Must be a list of provider names: "gemini", "qwen3", "hoppou", "chirp3_hd", "tiktok"

switchable_providers: ["gemini"]

# Qwen3 TTS server settings (only used when provider is "qwen3")

qwen3:

# URL of your Qwen3 TTS server

base_url: "http://localhost:7860"

# TTS mode: "voice_clone", "custom", or "voice_design"

mode: "voice_clone"

# Language for synthesis

language: "English"

# ── Voice Clone settings ──

# Use a preset reference voice (from server's preset list)

ref_preset: ""

# OR provide a local reference audio file + transcript

ref_audio: ""

ref_text: ""

# x-vector only (faster) vs full clone (more accurate)

xvec_only: true

# ── Custom Voice settings ──

# Speaker name (for "custom" mode, requires CustomVoice model)

speaker: ""

# ── Voice Design / Custom instruct ──

# Text description of desired voice (for "voice_design" or "custom" mode)

instruct: ""

# ── Generation parameters ──

chunk_size: 8 # Streaming chunk size (tokens per chunk)

temperature: 0.9

top_k: 50

repetition_penalty: 1.05

# Hoppou AI cloud TTS settings (only used when provider is "hoppou")

hoppou:

# API endpoint

api_url: "https://api.hoppou.ai/tts"

# API key (get one at https://hoppou.ai/api-portal/)

api_key: ""

# Voice preset name

voice: "alba"

# Model name

model: "tts-1"

# Google Cloud Chirp 3: HD TTS settings (only used when provider is "chirp3_hd")

# Uses the streaming API for low-latency synthesis with Google's latest HD voices.

# Requires google-cloud-texttospeech package.

chirp3_hd:

# API key for Google Cloud TTS (get from https://console.cloud.google.com/apis/credentials)

# Leave empty to use Application Default Credentials (ADC) instead

api_key: ""

# Backup keys for automatic rotation on rate limit (429) errors

# Free tier: ~500k chars/month. When exhausted, rotates to next key automatically.

backup_keys: []

# - "BACKUP_CLOUD_KEY_1"

# - "BACKUP_CLOUD_KEY_2"

# Voice name -- see https://docs.cloud.google.com/text-to-speech/docs/chirp3-hd

# Available: Achernar, Achird, Algenib, Algieba, Alnilam, Aoede, Autonoe,

# Callirrhoe, Charon, Despina, Enceladus, Erinome, Fenrir, Gacrux, Iapetus,

# Kore, Laomedeia, Leda, Orus, Puck, Pulcherrima, Rasalgethi, Sadachbia,

# Sadaltager, Schedar, Sulafat, Umbriel, Vindemiatrix, Zephyr, Zubenelgenubi

voice: "Kore"

# Language code (e.g. en-US, ja-JP, de-DE, fr-FR)

language_code: "en-US"

# Speaking rate (0.25 to 2.0, 1.0 = normal)

speaking_rate: 1.0

# TikTok TTS settings (only used when provider is "tiktok")

# Uses the free community Weilbyte TikTok TTS proxy (no auth needed).

# See: https://github.com/Weilbyte/tiktok-tts

tiktok:

# Voice code to use. Full list: https://github.com/oscie57/tiktok-voice/wiki/Voice-Codes

#

# Popular voices:

# en_us_001 English US Female 1 en_us_006 English US Male 1

# en_us_002 English US Female 2 en_us_007 English US Male 2

# en_us_009 English US Male 3 en_us_010 English US Male 4

# en_uk_001 English UK Male 1 en_uk_003 English UK Male 2

# en_au_001 English AU Female en_au_002 English AU Male

# en_male_narration Narrator en_male_funny Wacky

# en_female_emotional Peaceful en_male_cody Serious

#

# Disney voices:

# en_us_ghostface Ghost Face en_us_chewbacca Chewbacca

# en_us_c3po C3PO en_us_stitch Stitch

# en_us_rocket Rocket en_us_stormtrooper Stormtrooper

# en_female_madam_leota Madame Leota en_male_ghosthost Ghost Host

# en_male_pirate Pirate

#

# Singing voices:

# en_female_f08_salut_damour Alto en_male_m03_lobby Tenor

# en_male_m03_sunshine_soon Sunshine Soon

# en_female_f08_warmy_breeze Warmy Breeze

# en_female_ht_f08_glorious Glorious

# en_male_sing_funny_it_goes_up It Goes Up

# en_male_m2_xhxs_m03_silly Chipmunk

# en_female_ht_f08_wonderful_world Dramatic

#

# Other languages:

# fr_001 / fr_002 French de_001 / de_002 German

# es_002 Spanish es_mx_002 Spanish MX

# br_001 / br_003 / br_004 / br_005 Portuguese BR

# jp_001 / jp_003 / jp_005 / jp_006 Japanese

# kr_002 / kr_003 / kr_004 Korean

# id_001 Indonesian

voice: "en_us_001"

audio:

# Audio device indices - null uses system default

# Run `python -c "import pyaudio; p=pyaudio.PyAudio(); [print(i, p.get_device_info_by_index(i)['name']) for i in range(p.get_device_count())]"` to list devices

input_device: null

output_device: null

# Sample rates (don't change unless you know what you're doing)

send_sample_rate: 16000

receive_sample_rate: 24000

chunk_size: 1024

# Thinking sound effect -- plays a looping ambient sound while the AI is thinking

# or recalling memories. Fades in when thinking starts, fades out when done.

# Place your sound file (e.g. a subtle hum or chime loop) at the configured path.

thinking_sound:

enabled: false

on_thinking: true # Play during AI thinking (thought generation)

on_recall: true # Play during memory recall (recallMemories tool)

file: "sfx/thinking.wav" # Path to the sound file (wav/ogg/mp3)

volume: 30 # Volume 0-100

fade_in_ms: 500 # Fade-in duration in milliseconds

fade_out_ms: 800 # Fade-out duration in milliseconds

# Real-time voice pitch shifting -- allows the AI to shift its voice pitch

# up or down in semitones via the setVoicePitch tool, like a voice changer.

pitch_shift:

enabled: false # Set to true to allow pitch shifting

max_semitones: 12 # Maximum allowed shift in either direction (1-24)

vrchat:

# VRChat OSC settings

osc_ip: "127.0.0.1"

osc_send_port: 9000

osc_receive_port: 9001 # Port to receive avatar parameters from VRChat (velocity, grounded, etc.)

# Seconds to display each chatbox page before advancing

chatbox_page_delay: 3.0

# Shared limiter for every VRChat chatbox sender, including model text, music UI, idle banners, and tools.

# VRChat now allows short bursts instead of a flat timeout. Keep enabled unless you know your setup is exempt.

chatbox_rate_limiter:

enabled: true

capacity: 5

window_seconds: 5.0

safety_margin_seconds: 0.1

# Used only when the leaky bucket limiter is disabled.

legacy_min_interval_seconds: 1.27

# Idle chatbox banner -- shown in VRChat chatbox when the AI is idle.

# Displays a customizable banner with up to 3 lines of text, dividers,

# active session time, and current clock. Respects the 144-char chatbox limit.

idle_chatbox:

enabled: false

# Header text at the top of the banner

banner: "Gabriel AI"

# Character used for divider lines (also used in the live music gen display)

divider: "\u2500"

# Number of times the divider character is repeated (also used in the live music gen display)

divider_length: 14

# Up to 3 lines of text between dividers (empty strings are skipped)

lines:

- "VRChat AI Assistant"

- "Listening for voice"

- ""

# How often (seconds) to refresh the banner (updates clock/active time)

update_interval: 30

# ══════════════════════════════════════════════════════════

# VRChat API (for avatar switching)

# ══════════════════════════════════════════════════════════

# Required for searchAvatars and switchAvatar tools.

# Credentials are used to authenticate with the VRChat API.

# Session cookies are persisted to data/vrchat_cookies.json.

vrchat_api:

# VRChat account username (or email)

username: ""

# VRChat account password

password: ""

# TOTP secret for automatic 2FA (optional)

# This is the secret key from your authenticator app setup (NOT the 6-digit code).

# If set, 2FA is handled automatically. Requires: pip install pyotp

# Leave empty to skip auto-2FA (you'll need to manually provide codes).

totp_secret: ""

# Allow the model to edit your VRChat profile bio via updateStatus tool.

# When false (default), only status and statusDescription can be changed.

allow_bio_edit: false

# Group ID for the inviteToGroup tool. The model can invite people to this

# group when asked. Format: grp_00000000-0000-0000-0000-000000000000

# Leave empty to disable the inviteToGroup tool. You must have invite

# permissions in the group (usually group owner or roles with that power).

group_id: ""

music:

# Folder containing local music files (.mp3, .wav, .ogg, .flac)

music_dir: "sfx/music"

yolo:

# Master switch -- set to false to completely disable the player-following system

# (removes the tool declarations from Gemini and skips tracker init)

enabled: true

# Live YOLO stream + per-setting sliders are in the main WebUI under the Vision tab.

# YOLOv8n model auto-downloads here if missing

model_dir: "models/yolov8"

model_name: "yolov8n.pt"

face_tracker:

# Face tracking - smoothly looks at faces using yolov8n-face model

# When AI is speaking: locks onto the closest face

# When idle: randomly glances at visible faces every 5-10 seconds

enabled: false

wanderer:

# Autonomous wandering using depth estimation for obstacle avoidance

# AI can toggle this to freely explore VRChat maps while avoiding walls

enabled: false

# Depth model: "depth-anything-v2-small" (fast, ~50ms), "depth-anything-v2-base" (balanced),

# "dpt-large" (heavy, ~450ms)

model: "depth-anything-v2-small"

# Use FP16 on GPU (halves VRAM, ~2x faster, no quality loss)

fp16: true

# Remote depth server (offload inference to a dedicated GPU machine)

# See depth_server/ for the server setup

depth_server:

enabled: false

url: "http://192.168.1.x:8780"

api_key: "your-secret-key"

vision:

# Screen capture for Gemini Live vision

# Allows the AI to SEE what's on your screen

enabled: true

# Monitor index to capture (0 = all monitors combined, 1 = primary, 2 = secondary, etc.)

# Run `python -c "import mss; s=mss.mss(); [print(i, m) for i,m in enumerate(s.monitors)]"` to list

monitor: 1

# Seconds between screen captures (lower = more responsive but uses more tokens)

# Note: audio+video sessions are limited to 2 minutes, use context window compression

# For 3.1 models on free tier: auto-bumped to 2.0s if set below that (token optimization)

interval: 1.0

# Max resolution (images are scaled to fit this size, preserving aspect ratio)

# For 3.1 models: auto-capped at 768 to reduce payload size

max_size: 1024

# JPEG quality (1-100, higher = better quality but larger payload)

# For 3.1 models: auto-capped at 60 to reduce payload size

quality: 80

# Media resolution controls how many tokens each image costs in the Live API.

# Options: "low" (280 tokens), "medium" (560 tokens), "high" (1120 tokens)

# Auto-defaults to "low" for 3.1 models if not set (critical for free tier 65K TPM limit).

# For 2.5 models the default is unset (uses API default ~256 tokens with pan & scan).

# media_resolution: "low"

# Pause screen capture while AI is speaking or music is playing (saves tokens).

# Live music (music_gen) is excluded from this pause so the AI can still see reactions.

# Set to false if you want the AI to always see the screen even while talking.

pause_on_output: true

# Pause screen capture when nobody is interacting with the AI.

# Instead of fully stopping, vision slows down to idle_interval to save tokens

# while still keeping some awareness. Resumes normal speed when someone speaks

# or any task becomes active. Works with both auto and Silero VAD modes.

pause_on_idle: true

# Seconds between vision frames when idle. Only used when pause_on_idle is true.

# Lower = more awareness while idle but more tokens, higher = saves tokens.

idle_interval: 15.0

# ══════════════════════════════════════════════════════════

# Persistent Memory System

# ══════════════════════════════════════════════════════════

memory:

# Enable memory system

enabled: true

# Backend: "sqlite" (local file) or "mongo" (MongoDB Atlas)

backend: "sqlite"

# SQLite settings (used if backend is "sqlite")

sqlite_path: "data/gabriel_memories.sqlite"

# MongoDB settings (used if backend is "mongo")

# Set GABRIEL_MONGO_URI environment variable, or configure below

mongo_uri: ""

mongo_db: ""

mongo_collection: ""

# Memory TTL (time-to-live) settings

quick_note_ttl_hours: 6 # Quick notes auto-delete after 6 hours

short_term_ttl_days: 7 # Short-term memories auto-delete after 7 days

# Rate limiting for quick notes (prevents spam)

note_min_interval_seconds: 120 # Min seconds between quick notes

dedupe_window_seconds: 300 # Deduplication window

# How many recent memories to include in the system prompt

prompt_memory_count: 15

# RAG System (semantic memory recall via embeddings + vector search)

# When disabled, uses legacy keyword recall.

rag_enabled: false

# RAG provider: "gemini" (cloud, requires MongoDB) or "local" (ChromaDB + LM Studio)

# "gemini" - uses Gemini embedding API + MongoDB Atlas vector search

# "local" - uses local embedding model via LM Studio + ChromaDB vector DB (no cloud needed)

rag_provider: "gemini"

# -- Gemini RAG settings (rag_provider: "gemini") --

# Requires MongoDB backend. Gemini embedding API has 100 RPM limit on free tier.

embedding_model: "gemini-embedding-001" # Gemini embedding model

embedding_dims: 768 # Output dimensions (lower = faster, max 3072)

# -- Local RAG settings (rag_provider: "local") --

# Requires LM Studio running with an embedding model loaded.

# Works with any backend (sqlite or mongo).

# See README for setup instructions.

lm_studio_url: "http://localhost:1234" # LM Studio server URL

local_embedding_model: "text-embedding-embeddinggemma-300m-qat" # Model loaded in LM Studio

chroma_dir: "data/gabriel_chroma_db" # ChromaDB storage directory

# Vector similarity thresholds (per-provider, since local models score lower)

vector_min_score_gemini: 0.82 # Gemini embeddings: higher scores, stricter threshold

vector_min_score_local: 0.55 # Local embeddings (LM Studio): lower scores, looser threshold

# vector_min_score: 0.82 # (legacy) single threshold for both, overrides the active provider

# ══════════════════════════════════════════════════════════

# Avatar Emotion/Animation System

# ══════════════════════════════════════════════════════════

emotions:

# Enable emotion system

enabled: false

# VRChat Avatar ID (optional, for logging/tracking)

avatar_id: ""

# Default duration for non-looping animations (seconds)

default_duration: 3.0

# How often to switch between talking animations when speaking (seconds)

talking_switch_interval: 5.0

# Idle animation - plays when nobody is speaking for a while

idle_enabled: true # Enable/disable idle animation

idle_animation: "sad-idle" # Animation name to play when idle (must be defined below)

idle_timeout: 10 # Seconds of silence before idle animation starts

# Thinking animation - plays while AI is thinking or recalling memories

# Set to a defined animation name, or leave empty to disable

thinking_animation: "" # Animation name to play when thinking (must be defined below)

# ── Animations ──

# Each animation maps to its VRChat OSC parameter path

# looping: true = stays on until stopAnimation is called

# auto_talking: true = auto-triggered when AI speaks (not manually callable)

animations:

# Example emotion animation

# sad-idle:

# osc_path: "/avatar/parameters/YourParam/sad-idle"

# category: "emotion"

# looping: true

# Example talking animation (auto-triggered when AI speaks)

# talking-1:

# osc_path: "/avatar/parameters/YourParam/talking-1"

# category: "talking"

# looping: true

# auto_talking: true

# Example dance animation

# dance-1:

# osc_path: "/avatar/parameters/YourParam/dance-1"

# category: "dance"

# looping: true

# ══════════════════════════════════════════════════════════

# OBS Overlay (optional streaming overlay)

# ══════════════════════════════════════════════════════════

# Provides transparent browser source overlays for OBS/streaming.

# When disabled, overlay routes and music broadcast are inactive.

obs:

enabled: false

# ══════════════════════════════════════════════════════════

# Lyria RealTime Music Generation (experimental)

# ══════════════════════════════════════════════════════════

# AI-powered real-time instrumental music generation using Google's Lyria RealTime.

# When enabled, the AI can generate and perform live instrumental music that can

# be steered in real-time (change style, tempo, key, density, brightness, etc.).

# Uses the same Gemini API key. Requires v1alpha API access.

# Output: stereo 48kHz 16-bit PCM via a dedicated audio stream.

music_gen:

# Master switch - set to false to completely disable music generation

# (removes tool declarations from Gemini and skips module init)

enabled: false

# Default BPM (60-200). Leave null to let the model decide based on prompts.

default_bpm: 120

# Default playback volume (0-200, 100 = normal)

volume: 80

# Temperature (0.0-3.0, default 1.1)

temperature: 1.1

# Prompt guidance (0.0-6.0, default 4.0). Higher = stricter prompt following.

guidance: 4.0

# Mute bass by default (solo instrument feel)

mute_bass: true

# Mute drums by default (solo instrument feel)

mute_drums: true

# ══════════════════════════════════════════════════════════

# Web Search (Jina Reader API)

# ══════════════════════════════════════════════════════════

# Provides webSearch and readWebpage tools via Jina Reader API.

# s.jina.ai for web search, r.jina.ai for URL content extraction.

# Only active on Gemini 3.1 models (2.5 models use built-in Google Search instead).

web_search:

enabled: false

# Optional Jina API key for higher rate limits

# Free without key: 20 RPM. With key: 100-1000 RPM.

# Get a free key at https://jina.ai/reader/

jina_api_key: ""

# ══════════════════════════════════════════════════════════

# Discord Selfbot Integration

# ══════════════════════════════════════════════════════════

# Uses discord.py-self for a selfbot that has its own Gemini Live

# session and can respond to messages, relay info to VRChat, etc.

# Configuration lives in discord_bot/config.yml (separate file).

discord_bot:

enabled: false

# ══════════════════════════════════════════════════════════

# Social Server (AI-to-AI Messaging)

# ══════════════════════════════════════════════════════════

# Connect to a ProjectGabriel Social Server to message other people,

# manage friends, and see who's online. Run the server from social_server/.

# See social_server/README.md for setup instructions.

social:

enabled: false

# URL of the social server

server_url: "http://localhost:3000"

# Your API key (must match a key in the social server's config.yml)

# Leave blank if the server is in open mode

api_key: "YOUR_SOCIAL_API_KEY"

# Password for open mode servers (used to register/login without an API key)

# Leave blank if using API key auth

password: ""

# Your AI's username (must match the username for your key in the server config)

# In open mode, this is the username your AI will register with

username: "Gabriel"

# Short description of your AI shown to other users

description: "A VRChat AI companion"

# If true, your AI appears offline to others (won't show in online lists)

# You can still send/receive messages, just won't appear online

appear_offline: false

# How often to send heartbeat pings (seconds)

heartbeat_interval: 30

# How often to poll for new messages when WebSocket is unavailable (seconds)

message_check_interval: 60

# Seconds to wait after receiving a message before prompting a reply.

# Messages arrive as context (turn_complete=false). After this delay

# with no new messages, the AI is nudged to consider replying (turn_complete=true).

idle_reply_delay: 300

# ══════════════════════════════════════════════════════════

# Plugins

# ══════════════════════════════════════════════════════════

# Drop-in plugin system. Anything in ./plugins/<name>/ with a plugin.yml

# manifest gets auto-loaded on startup. See plugins/README.md for the

# full author guide. Plugins can register Gemini tools, TTS providers,

# STT providers, and subscribe to lifecycle events (startup, shutdown,

# message_in, message_out).

#

# Per-plugin enable/disable lives in config/tools.yml under the `plugins:`

# block, alongside per-tool toggles. This section only holds runtime

# settings the plugins read via ctx.plugin_config(...).

plugins:

# Master toggle for the whole plugin system. Set false to skip loading

# any plugins at all (useful for debugging if a plugin is misbehaving).

enabled: true

# Trust mode. When false (default) plugins get a sandboxed view of the

# host config that hides secrets (gemini api_key, vrchat password,

# mongo connection string, discord token, etc). Plugins should store

# their own settings under plugins.<name>.* and read them through

# ctx.plugin_config(), not by reaching into ctx.config.api_key.

#

# Flip this to true if you have an older plugin that needs raw access

# to those, like the diary plugin which uses the main gemini api_key

# for its background sub-agent. Only do this for plugins you trust,

# any loaded plugin will be able to read every secret in this file.

trusted: false

# Per-tool and per-plugin enable toggles live in config/tools.yml.

# Run the configurator or edit that file directly to hide individual

# tools from Gemini.