Skip to content

Latest commit

 

History

History
121 lines (93 loc) · 5.79 KB

README.md

File metadata and controls

121 lines (93 loc) · 5.79 KB

PlaylistRAGulator

Python Docker Shell

Flow Diagram

About

PlaylistRAGulator is a RAG AI app that matches your vibe. Drop in some text or an image, and it’ll curate a playlist that perfectly fits the mood. Effortless music discovery, one vibe at a time.

Local Execution

PlaylistRAGulator is fully containerized using Docker, ensuring a seamless setup with minimal effort. All models run locally through Ollama, providing a secure and efficient environment. Once the Docker container is built, no additional installation is required—everything you need is bundled inside the container.

How to Use

1. Docker Setup

  • Navigate to the dockerfile folder and build the Docker image: 
    docker build -t playlistragulator .

2. Download Models

  • Run the following command to download the required models:

⚠️ Note: Update file paths to match your local PlaylistRAGulator folder.

docker run --rm \
  -v /path/to/PlaylistRAGulator/ollama_models:/root/.ollama/models \
  -v /path/to/PlaylistRAGulator:/home \
  playlistragulator bash pull_models.sh

3. Generate Playlist

  • Unzip the database: 
    unzip data/chroma_db.zip
  • Run the playlist generator:

⚠️ Note: Update file paths to match your local PlaylistRAGulator folder.

docker run --rm \
  -v /path/to/PlaylistRAGulator/ollama_models:/root/.ollama/models \
  -v /path/to/PlaylistRAGulator:/home \
  playlistragulator bash make_playlist.sh

Input & Output

Input

The inputs can be found in the folder data/user_inputs.

  • Text Prompt: A description of the desired mood, vibe, or theme (e.g., "Chill sunset vibes" or "Energetic workout beats").
  • Image: Upload an image that represents the mood or location you want to capture in the playlist.
  • Configuration File: Located at data/user_inputs/config.json, this file controls model selection, file paths, and generation settings.

Config File Explanation

  • embedmodel: Name of the embedding model used for text analysis (e.g., "mxbai-embed-large").
  • vismodel: Visual model used to analyze images (e.g., "llava:7b").
  • chroma_db_path: Path to the Chroma database used for song embeddings.
  • nb_songs: Number of songs to include in the generated playlist.
  • dataset_in_path: Path to the input dataset containing music metadata.
  • use_img: Boolean flag to enable or disable image input (true to use images).
  • img_path: Path to the input image for mood detection.
  • img_small_path: Path to the resized version of the input image.
  • img_w: Width of the resized image.
  • img_h: Height of the resized image.
  • img_desc_path: Path to the generated image description text file.
  • use_adj: Boolean flag to use additional adjectives for mood enhancement.
  • adj_path: Path to a CSV file containing mood adjectives.
  • output_path: Path where the generated playlist will be saved.
  • nb_row_embed: Number of rows to embed from the dataset for analysis.
  • songs_path: Path to the CSV file containing song metadata.
  • features_path: Path to the CSV file containing acoustic features of songs.

Output

  • Curated Playlist: A playlist of songs tailored to match the mood or theme based on the input.
  • Playlist File: The playlist is saved as data/outputs/playlist.csv, where each row represents a selected song.
  • Processed Image & Description: In data/outputs/inter/, you will find the resized input image and the text description generated by the model.
  • File Format: The playlist can be output as a CSV file, a text file with song titles.

How It Works

Flow Diagram

  1. Image Analysis: If an image is provided, it is processed by the vision model (llava:7b) to generate a descriptive text that captures the atmosphere and mood.
  2. Prompt Generation: This description is combined with a user-provided list of adjectives (from data/user_inputs/adjectives.csv) to create a detailed prompt.
  3. Music Matching: The prompt is used to query the vector database (data/chroma_db) that contains embeddings of songs, selecting tracks that best match the mood.

Here you can find an example how the data flows into the pipeline.

Flow Diagram

Data & Customizing the Vector Database

The vector database used by PlaylistRAGulator is built from the Musicoset Dataset. To customize the database:

  1. Download the Dataset: Place the dataset in the data folder.

  2. Run Preprocessing: Execute the following command to preprocess the data and create the vector database:

⚠️ Note: Update file paths to match your local

docker run --rm \
  -v /Users/lorenzosalmina/Desktop/PlaylistRAGulator/ollama_models:/root/.ollama/models \
  -v /Users/lorenzosalmina/Desktop/PlaylistRAGulator:/home \
  playlistragulator bash setup.sh

  1. What Happens:

    • The preprocessing.py script generates musicoset.csv with song titles and features.
    • The make_vector_db.py script embeds each row of musicoset.csv to create the vector database.
    • You can control how many rows are embedded by adjusting the "nb_row_embed" field in data/user_inputs/config.json.

  2. Result:

    • A new folder data/chroma_db is created, containing the embedded songs and features.
    • To customize the dataset or features, modify the preprocessing.py script.

Enjoy discovering music that matches your vibe with PlaylistRAGulator! 🎶