Gemma YTMusic Assistant

A natural-language YouTube Music assistant powered by FunctionGemma — a 270M-parameter Gemma 3 model fine-tuned for structured function calling. Type plain English; the model dispatches to the right tool; your browser opens (or a list is printed).

Overview

Setup

Requirements: Python with torch, transformers, huggingface_hub, google-auth-oauthlib, sentence-transformers, questionary

# Create and activate virtual environment
python -m venv .venv
.venv\Scripts\activate        # Windows
source .venv/Scripts/activate # Git Bash / WSL

# Install dependencies
pip install torch transformers huggingface_hub google-auth-oauthlib google-api-python-client sentence-transformers questionary

# Authenticate with Hugging Face (one-time, token stored in GemmaToken.txt)
python hf_login.py

YouTube API setup:

1. Create a project in Google Cloud Console
2. Enable YouTube Data API v3
3. Create an OAuth 2.0 Client ID (Desktop app) and download it as client_secrets.json
4. On first run, a browser tab opens for Google login — token.json is written and auto-refreshed

Running

python main.py              # minimal weather demo (no API keys needed)
python ytmusic_assistant.py # interactive YouTube Music assistant

Available Tools

How It Works

FunctionGemma operates on a fixed 4-stage cycle:

1. Build messages — processor.apply_chat_template(messages, tools=[...]) encodes the user prompt with tool schemas
2. Model inference — model.generate() emits a structured function call:

   <start_function_call>call:function_name{param:<escape>value<escape>}<end_function_call>
   

3. Execute tool — parse with regex, look up in TOOL_MAP, call the Python function
4. Return result — print output or open browser

Critical: The role: "developer" system message "You are a model that can do function calling with the following functions" must be present — it activates function-calling behavior.

Recommendation Engine

Recommendations use a scoring formula combining semantic similarity and learned taste:

score = alpha * cosine_sim(prompt_embedding, title_embedding)
      + beta  * mean(preference_profile[dim][label])

Video titles are labeled across 5 dimensions (content type, genre, mood, use case, language) using all-MiniLM-L6-v2 embeddings. Your preference profile (data/preferences.json) updates via EMA each time you select recommended videos.

Local Data Files

Model

google/functiongemma-270m-it — 270M parameter Gemma 3 model fine-tuned exclusively for single-turn function calling. It does not produce free-form text answers.

Limitations: single-turn only, no chaining, no multi-step reasoning.