通过 API 使用 Nova 的文档理解功能

下面是一个简化版的 Python 示例，说明了如何使用 Amazon Nova 进行文档 QA（问答）或分析。我们将使用 AWS Bedrock API（通过 Boto3 SDK）发送一份 PDF 文档和一个供模型回答的问题。

            
import base64
import base64
import json
import boto3

# Initialize Bedrock runtime client (adjust region as needed)
client = boto3.client("bedrock-runtime", region_name="us-east-1")

MODEL_ID = "us.amazon.nova-lite-v1:5"  # using Nova Lite model in this example

# Read the document file (PDF) in binary mode
with open("my_document.pdf", "rb") as file:
    doc_bytes = file.read()

# Construct the conversation messages with document + question
messages = [
    {
        "role": "user",
        "content": [
            {
                "document": {
                    "format": "pdf",
                    "name": "Document1",  # neutral name for the document
                    "source": {
                        "bytes": doc_bytes  # embedding the PDF content directly
                    }
                }
            },
            {
                "text": "Here is a question about the document: ... (your question) ... ?"
            }
        ]
    }
]

# Set inference parameters (optional)
inf_params = {"maxTokens": 4000, "topP": 0.1, "temperature": 0.3}

# Invoke the model
response = client.converse(modelId=MODEL_ID, messages=messages, inferenceConfig=inf_params)

# Extract and print the answer
answer_text = response["output"]["message"]["content"][0]["text"]
print(answer_text)

如果输入文件很大（超过 25 MB 的直接上传限制）或文件很多，可将其存储在 Amazon S3 中并进行引用。这就避免了在请求中发送原始字节。使用 S3 时，需确保 Bedrock 服务有权访问存储桶/对象。例如，为了在 S3 中引用 PDF，文档源将使用“S3 位置”而不是“字节”，如下所示：

messages = [
    {
        "role": "user",
        "content": [
            {
                "document": {
                    "format": "pdf",
                    "name": "Report2023",
                    "source": {
                        "s3Location": {
                            "uri": "s3://your-bucket/path/to/document1.pdf",
                            "bucketOwner": "123456789012"
                        }
                    }
                }
            },
            {
                "text": "Summarize the key findings from the Q3 2023 report."
            }
        ]
    }
]