Invoking Amazon Nova understanding models

이 페이지는 귀하의 언어로 번역되지 않았습니다. 번역 요청

Amazon Nova Multimodal understanding models are available for use for inferencing through the Invoke API (InvokeModel, InvokeModelWithResponseStream) and the Converse API (Converse and ConverseStream). To create conversational applications see Carry out a conversation with the converse API operations. Both of the API methods (Invoke and Converse) follow a very similar request pattern as detailed below

Key Differences between the Invoke API and Converse API are as follows:

• Inference parameters like topK are not supported in Converse and need to be passed in additionalModelRequestFields, while in the Invoke API it can be passed directly in the inference parameters.

• Document Support is limited to only Converse API and is not supported in Invoke API.

• Response parsing formats are different between the Invoke API and Convserse API constructs.

• Response streaming is different between ConverseStream and InvokeModelWithStreaming.

Topics

• Messages API overview

• Utilizing the system prompt

• Using the Converse API

• Using the Invoke API

• Complete request schema

Messages API overview

The Amazon Nova Understanding models utilize the Messages API, which enables the submission of structured lists of input messages containing text, images, videos, and documents. The model then generates the next message in the conversation. The Messages API supports both single queries and stateless multi-turn conversations, allowing for the creation of chatbots and virtual assistant applications. The API manages the conversational exchanges between a user and an Amazon Nova model (assistant).

Amazon Nova models are trained to operate on alternating user and assistant conversational turns. When creating a new message, you specify the prior conversational turns with the messages parameter. The model then generates the next messages in the conversation.

Each input message must be an object containing a role and content. Users may specify a single message with the user role, or include multiple messages with both user and assistant roles. However, the first message must always use the user role. If the technique of prefilling the response from Amazon Nova is employed (by including a final message with the assistant role), the model will continue its response from the provided content. This approach will still result in a response with the assistant role.

The following represents a single user message:

[{
  "role": "user",
  "content": [{"text":"Hello, Nova"}]
}]

Here is an example with multiple conversational turns:

[
  {"role": "user", "content": [{"text": "Hello there."}]},
  {"role": "assistant", "content": [{"text": "Hi, I'm Chatbot trained to answer your questions. How can I help you?"}]},
  {"role": "user", "content": [{"text": "Can you explain LLMs in plain English?"}]}
]

Here is an example with a partially-filled response from Amazon Nova:

[
  {"role": "user", "content": [{"text":"Please describe yourself using only JSON"}]},
  {"role": "assistant", "content": [{"text":"Here is my JSON description:\n{"}]}
]

For information about creating prompts for Amazon Nova models, see Text understanding prompting best practices.

Utilizing the system prompt

You can include a system prompt in the request. A system prompt lets you provide context and instructions to Amazon Nova, such as specifying a particular goal or role. Specify a system prompt in the system field, as shown in the following example:

[
   {"text": "You are an expert SaS analyst......"}
]