bedrock-kit 0.1.0

Small, opinionated AWS Bedrock client wrapper: adaptive throttle, cache-aware cost tracking, and structured-output parse-and-repair. Single-cloud, single-purpose.

bedrock-kit

Small, opinionated AWS Bedrock client wrapper.

Every Bedrock production team rebuilds the same three things: adaptive throttle for ThrottlingException, per-call cost tracking that handles cache-read tokens, and structured-output parsing-with-repair. bedrock-kit ships those, and nothing else. Single-cloud, single-purpose. No proxy server. Wraps the official boto3 client; you can inject a fake for testing.

Install

pip install "bedrock-kit[boto]"
# optional
pip install "bedrock-kit[boto,pandas]"  # adds CostLedger.to_pandas()

Quickstart

from bedrock_kit import BedrockClient, AdaptiveThrottle, CostLedger

ledger = CostLedger()
client = BedrockClient(
    region="us-east-1",
    retry=AdaptiveThrottle(max_attempts=8, base_delay=1.0, max_delay=30.0),
    cost_ledger=ledger,
)

resp = client.invoke(
    model_id="anthropic.claude-sonnet-4-5",
    messages=[{"role": "user", "content": "Hello"}],
    system="You are concise.",
    max_tokens=512,
)

resp.text                       # the model's reply
resp.usage.input_tokens         # 12
resp.usage.cache_read_tokens    # 0
resp.cost_usd                   # 0.000176
resp.cache_hit                  # False
ledger.total_usd                # accumulates across all calls
ledger.by_model                 # {"anthropic.claude-sonnet-4-5": 0.000176}

Structured output with repair

from pydantic import BaseModel
from bedrock_kit import JsonSchema

class Sentiment(BaseModel):
    label: str
    confidence: float

resp = client.invoke(
    model_id="anthropic.claude-sonnet-4-5",
    messages=[{"role": "user", "content": "Classify: 'this is great!'"}],
    response_schema=JsonSchema(Sentiment, max_repair_attempts=2),
)
resp.parsed                  # Sentiment(label="positive", confidence=0.95)

If the model returns invalid JSON, bedrock-kit first does a light local repair pass (strip markdown fences, trailing commas, surrounding prose). If that still fails, it asks the model to fix its own output, up to max_repair_attempts times.

Adaptive throttle

throttle = AdaptiveThrottle(
    max_attempts=8,    # total attempts (incl. first)
    base_delay=1.0,    # seconds
    max_delay=30.0,
    jitter=True,       # full-jitter (uniform(0, capped_delay))
)

Retries on Bedrock throttle codes: ThrottlingException, TooManyRequestsException, ServiceUnavailableException, ProvisionedThroughputExceededException, ModelTimeoutException. Does not retry validation, auth, or model-not-found errors - those are your bugs to fix, not transient.

Cost tracking

CostLedger ships pricing for popular Anthropic Bedrock models. Override or extend with pricing=:

from bedrock_kit import CostLedger, Pricing

ledger = CostLedger(
    pricing={
        "amazon.nova-pro-v1:0": Pricing(
            input=0.8, output=3.2, cache_read=0.2, cache_write=1.0
        ),
    },
    strict=True,  # raise PricingNotFoundError on unknown models
)

Default pricing is best-effort and dated; verify against aws.amazon.com/bedrock/pricing before using these numbers for billing.

Why not LiteLLM?

LiteLLM is great if you need cross-provider routing. bedrock-kit is for the case where you've already decided on Bedrock, you don't want a 46k-LOC multi-provider abstraction, and you want a small surface a security team can audit. We don't proxy, don't include a server, don't ship a UI. We're < 1000 LOC of Python.

What it explicitly does NOT do

• No multi-provider routing
• No proxy server, no UI
• No prompt management
• No agent loop
• No image generation
• No SageMaker, Bedrock Agents, or Knowledge Bases SDK wrapping
• No streaming or cancellation yet (planned for v0.2)
• No OpenTelemetry emission yet (planned for v0.2)

Testing without AWS

The default BedrockClient makes a real boto3 client. For tests, inject a fake that quacks like one:

from bedrock_kit import BedrockClient, AdaptiveThrottle

class FakeClient:
    def converse(self, **kwargs):
        return {"output": {"message": {"content": [{"text": "stub"}]}},
                "stopReason": "end_turn",
                "usage": {"inputTokens": 1, "outputTokens": 1}}

client = BedrockClient(client=FakeClient(), retry=AdaptiveThrottle(sleep_fn=lambda _: None))

Status

v0.1 - alpha. Public API may change before v1.0. Issues and PRs welcome.