dialog 1.1.0

Manipulate and visualize LLMs conversations

💬 Dialog

Library to manipulate and display conversations.

✨ Try it on Colab: https://colab.research.google.com/github/google-deepmind/dialog/blob/main/colab/dialog.ipynb ✨

Features

• Create and manipulate conversations with minimal boilerplate

  conv = dialog.Conversation(
      dialog.User('What is this image ?\n', dialog.Image(data)),
      dialog.Model('This image represent a cat'),
      dialog.User('Thank you.'),
  )
  

• Round-trip conversions between:

  • Text: conv.as_text()
  • Tokens: conv.as_batch()

• Pretty Colab display

  • For conversation:
  • For text:

• Manipulations:

  • conv += dialog.User(): Append the next turn
  • conv += [dialog.User(), dialog.Model()]: Append multiple turns.
  • len(conv)
  • list(conv)
  • conv[-1][-1]: Slicing (last chunk of the last turn)

Thinking

• dialog.Think(): In the System instruction
• dialog.Thought('Model thoughts...'): In the model answer

conv = dialog.Conversation(
    dialog.System(dialog.Think()),
    dialog.User('Hello'),
    dialog.Model(
        dialog.Thought('I need to greet the user'),
        'Hello! What can I do for you?',
    )
)

Function calling

Function calling is done through:

• dialog.Tool: Tool definition in the system instruction
• dialog.ToolCall
• dialog.ToolResponse

conv = dialog.Conversation(
    dialog.System(
        dialog.Tool(tool0),
        dialog.Tool(tool1),
    ),
    dialog.User('Turn off the light in my bedroom'),
    dialog.Model(
        dialog.ToolCall(call0),
        dialog.ToolResponse(response0),
        'Lights have been turned off. Good night.',
    )
)

Multi-modalities

• dialog.Image: Image modality
• dialog.Audio: Audio modality

Modalities supports anything which can be interpreted as image/audio, including urls, paths, numpy array,...

dialog.User(
    'Describe those images:\n\n',
    dialog.Image(np.zeros((256, 256, 3), np.uint8)),
    dialog.Image('https://example.org/img.png'),
    dialog.Image('/path/to/my_img.jpg'),
),