agentchat.contrib.capabilities.vision_capability
VisionCapability
class VisionCapability(AgentCapability)
We can add vision capability to regular ConversableAgent, even if the agent does not have the multimodal capability, such as GPT-3.5-turbo agent, Llama, Orca, or Mistral agents. This vision capability will invoke a LMM client to describe the image (captioning) before sending the information to the agent's actual client.
The vision capability will hook to the ConversableAgent's process_last_received_message
.
Some technical details: When the agent (who has the vision capability) received an message, it will:
- _process_received_message: a. _append_oai_message
- generate_reply: if the agent is a MultimodalAgent, it will also use the image tag. a. hook process_last_received_message (NOTE: this is where the vision capability will be hooked to.) b. hook process_all_messages_before_reply
- send: a. hook process_message_before_send b. _append_oai_message
__init__
def __init__(lmm_config: Dict,
description_prompt: Optional[str] = DEFAULT_DESCRIPTION_PROMPT,
custom_caption_func: Callable = None) -> None
Initializes a new instance, setting up the configuration for interacting with a Language Multimodal (LMM) client and specifying optional parameters for image description and captioning.
Arguments:
lmm_config
Dict - Configuration for the LMM client, which is used to call the LMM service for describing the image. This must be a dictionary containing the necessary configuration parameters. Iflmm_config
is False or an empty dictionary, it is considered invalid, and initialization will assert.description_prompt
Optional[str], optional - The prompt to use for generating descriptions of the image. This parameter allows customization of the prompt passed to the LMM service. Defaults toDEFAULT_DESCRIPTION_PROMPT
if not provided.custom_caption_func
Callable, optional - A callable that, if provided, will be used to generate captions for images. This allows for custom captioning logic outside of the standard LMM service interaction. The callable should take three parameters as input:- an image URL (or local location)
- image_data (a PIL image)
- lmm_client (to call remote LMM)
and then return a description (as string).
If not provided, captioning will rely on the LMM client configured via
lmm_config
. If provided, we will not run the default self._get_image_caption method.
Raises:
AssertionError
- If neither a validlmm_config
nor acustom_caption_func
is provided, an AssertionError is raised to indicate that the Vision Capability requires one of these to be valid for operation.
process_last_received_message
def process_last_received_message(content: Union[str, List[dict]]) -> str
Processes the last received message content by normalizing and augmenting it with descriptions of any included images. The function supports input content as either a string or a list of dictionaries, where each dictionary represents a content item (e.g., text, image). If the content contains image URLs, it fetches the image data, generates a caption for each image, and inserts the caption into the augmented content.
The function aims to transform the content into a format compatible with GPT-4V multimodal inputs, specifically by formatting strings into PIL-compatible images if needed and appending text descriptions for images. This allows for a more accessible presentation of the content, especially in contexts where images cannot be displayed directly.
Arguments:
content
Union[str, List[dict]] - The last received message content, which can be a plain text string or a list of dictionaries representing different types of content items (e.g., text, image_url).
Returns:
str
- The augmented message content
Raises:
AssertionError
- If an item in the content list is not a dictionary.
Examples:
Assuming self._get_image_caption(img_data)
returns
"A beautiful sunset over the mountains" for the image.
-
Input as String: content = "Check out this cool photo!"
-
Output
- "Check out this cool photo!" (Content is a string without an image, remains unchanged.)- Input as String, with image location: content = "What's weather in this cool photo: <img http://example.com/photo.jpg>"
-
Output
- "What's weather in this cool photo: <img http://example.com/photo.jpg> in case you can not see, the caption of this image is: A beautiful sunset over the mountains " (Caption added after the image)- Input as List with Text Only: content = [{"type": "text", "text": "Here's an interesting fact."}]
-
Output
- "Here's an interesting fact." (No images in the content, it remains unchanged.)- Input as List with Image URL: content = [
-
{"type"
- "text", "text": "What's weather in this cool photo:"}, -
{"type"
- "image_url", "image_url": {"url": "http://example.com/photo.jpg"}} ] -
Output
- "What's weather in this cool photo: <img http://example.com/photo.jpg> in case you can not see, the caption of this image is: A beautiful sunset over the mountains " (Caption added after the image)