Bedrock Files and Batches - ModRouter

Endpoints

Use two clients when you need both runtime and S3-style file operations:

import os
import boto3

base_url = os.environ["MODROUTER_BASE_URL"].rstrip("/")

bedrock = boto3.client(
    "bedrock-runtime",
    endpoint_url=f"{base_url}/bedrock",
    region_name="us-east-1",
    aws_access_key_id="placeholder-access-key",
    aws_secret_access_key="placeholder-secret-key",
)

s3 = boto3.client(
    "s3",
    endpoint_url=f"{base_url}/bedrock/files",
    region_name="us-east-1",
    aws_access_key_id="placeholder-access-key",
    aws_secret_access_key="placeholder-secret-key",
)

S3-Compatible File Routes

ModRouter registers file routes under /bedrock/files:

Operation	ModRouter route
put object	`PUT /bedrock/files/{bucket}/{key:*}`
get object	`GET /bedrock/files/{bucket}/{key:*}`
delete object	`DELETE /bedrock/files/{bucket}/{key:*}`
head object	`HEAD /bedrock/files/{bucket}/{key:*}`
list bucket	`GET /bedrock/files/{bucket}`

Example:

s3.put_object(
    Bucket="modrouter-batch-input",
    Key="requests.jsonl",
    Body=b'{"recordId":"1","modelInput":{"messages":[{"role":"user","content":[{"text":"Hello"}]}]}}\n',
)

Batch Jobs

ModRouter registers Bedrock model invocation job routes:

Operation	ModRouter route
create	`POST /bedrock/model-invocation-job`
list	`GET /bedrock/model-invocation-jobs`
retrieve	`GET /bedrock/model-invocation-job/{job_arn}`
stop	`POST /bedrock/model-invocation-job/{job_arn}/stop`

Example shape:

job = bedrock.create_model_invocation_job(
    jobName="modrouter-batch-example",
    roleArn="arn:aws:iam::123456789012:role/BedrockBatchRole",
    modelId="bedrock/anthropic.claude-3-5-sonnet-20240620-v1:0",
    inputDataConfig={
        "s3InputDataConfig": {
            "s3Uri": "s3://modrouter-batch-input/requests.jsonl"
        }
    },
    outputDataConfig={
        "s3OutputDataConfig": {
            "s3Uri": "s3://modrouter-batch-output/"
        }
    },
)

Provider Notes

Bedrock-native batch jobs are S3-backed.
OpenAI batches use /openai/v1/batches and JSONL input files.
Anthropic message batches use /anthropic/v1/messages/batches with inline requests.