lmux-aws-bedrock

AWS Bedrock provider for lmux. Uses boto3, aiobotocore, and the Converse API.

Supports chat completions, streaming, and embeddings.

Part of the lmux ecosystem: standardized interface, cost tracking on every response, and registry-based routing across providers.

Optional Extras

• lmux-aws-bedrock[async]: async support via aiobotocore

Auth

Uses boto3's default credential chain (env vars, AWS config, instance metadata). No extra setup needed if your AWS credentials are already configured.

from lmux_aws_bedrock import BedrockProvider

provider = BedrockProvider()

# Or specify a region
provider = BedrockProvider(region="us-east-1")

For explicit session configuration:

from lmux_aws_bedrock import BedrockSessionAuthProvider

provider = BedrockProvider(auth=BedrockSessionAuthProvider(profile_name="my-profile"))

Usage

Chat

from lmux import UserMessage

response = provider.chat("anthropic.claude-sonnet-4-20250514-v1:0", [UserMessage(content="Hello")])
print(response.content)
print(response.cost)

Streaming

for chunk in provider.chat_stream("anthropic.claude-sonnet-4-20250514-v1:0", [UserMessage(content="Hello")]):
    if chunk.delta:
        print(chunk.delta, end="")

Embeddings

response = provider.embed("amazon.titan-embed-text-v2:0", "Hello")
print(response.embeddings)

Async

Requires the [async] extra. All methods have async variants: achat, achat_stream, aembed.

Bedrock also supports lmux response_format, mapped to Converse outputConfig.textFormat.

Registry

Use with the lmux registry to route across multiple providers:

from lmux import Registry

registry = Registry()
registry.register("bedrock", provider)
response = registry.chat("bedrock/anthropic.claude-sonnet-4-20250514-v1:0", messages)

Provider Params

from lmux_aws_bedrock import BedrockParams, GuardrailConfig

response = provider.chat(
    "anthropic.claude-sonnet-4-20250514-v1:0",
    messages,
    provider_params=BedrockParams(
        guardrail_config=GuardrailConfig(
            guardrail_identifier="my-guardrail",
            guardrail_version="1",
        ),
    ),
)

Prompt Caching

Place CachePointContent parts in UserMessage content to emit Converse cachePoint blocks marking the end of a stable prompt prefix. A cache point with no preceding block in its message is placed after whatever came before it (the prior message, or the system blocks). Markers with nothing cacheable before them are dropped, and adjacent duplicates are coalesced — the first marker wins.

from lmux import CachePointContent, TextContent, UserMessage

messages = [
    UserMessage(content=[TextContent(text=big_stable_context), CachePointContent()]),
    UserMessage(content="What changed since yesterday?"),
]

Cache points are emitted for whatever model the request targets; models without prompt-caching support reject them at request validation. Cache reads/writes are reported on response.usage (cache_read_tokens, cache_creation_tokens, and the per-TTL cache_creation_tokens_by_ttl breakdown from cacheDetails) and priced into response.cost, including per-TTL write rates where the pricing data carries them.

Constructor Options

BedrockProvider(
    auth=...,          # AuthProvider, default: BedrockEnvAuthProvider()
    region=...,        # AWS region
    endpoint_url=...,  # Custom endpoint URL
)