digline-bedrock

An Amazon Bedrock target and judges for digline, on the Converse API: a prompt file goes in, a priced Response comes out.

pip install digline-bedrock

Requires Python 3.12+, like digline itself.

Quickstart

No credential argument exists: the AWS chain — environment, profile, IAM role, instance metadata — is boto3's job, and this package never reads it.

from digline_bedrock import BedrockTarget

target = BedrockTarget(
    "prompts/answer.md",
    model="eu.anthropic.claude-sonnet-4-20250514-v1:0",
    max_tokens=1024,
)

model takes a model id or an inference profile id. The region is resolved when the target is built — from region= if you pass one, otherwise from the client the chain produced — and the price list follows from it:

from digline_bedrock import BedrockTarget

target = BedrockTarget(
    "prompts/answer.md",
    model="us.anthropic.claude-haiku-4-5-20251001-v1:0",
    max_tokens=1024,
    region="us-east-1",
)

target.region is read-only, and that is the point: what was priced is what was called. A missing region fails there, when the target is built, not on case thirty-seven with thirty-six paid calls behind it.

The judges run in the same account

A plugin is a target and a judge (ADR 0004), which on Bedrock is usually the whole reason the model is there: what a judge is sent is the model's output, and it stays inside the same account, the same region and the same IAM role.

from digline.core import LlmRubric
from digline_bedrock import BedrockJudge

judge = BedrockJudge(model="eu.anthropic.claude-haiku-4-5-20251001-v1:0")
rubric = LlmRubric(
    rubric="The answer is one sentence and cites the passage it came from.",
    judge=judge,
    threshold=0.8,
    tolerance=0.05,
)

from digline.core import Faithfulness
from digline_bedrock import BedrockClaimJudge

faithful = Faithfulness(
    judge=BedrockClaimJudge(model="eu.anthropic.claude-haiku-4-5-20251001-v1:0"),
    threshold=0.9,
    tolerance=0.05,
)

Converse has no structured-output mode, so the reply shape is asked for in the system prompt and read back leniently — a fenced block or a sentence in front of the object both parse. What judging cost is counted on the judge and never reset:

from digline_bedrock import BedrockJudge

judge = BedrockJudge(model="eu.anthropic.claude-haiku-4-5-20251001-v1:0")
print(f"{judge.calls} judgements, {judge.spent_usd:.4f} USD, {judge.latency_ms:.0f} ms")

Prices, and what is not priced

bedrock_pricing(region) is seeded for us-east-1, us-west-2, eu-west-1, eu-central-1 and eu-west-3, with the Anthropic models AWS sells in each, read from AWS's Price List API on the date PRICES_READ_ON carries. A model is priced only under the ids that reach it there: its us. or eu. profile at the Standard price, global. at the global one (about 10% less from Haiku 4.5 on), and its bare id only where AWS takes it. SEEDED_PRICES is the whole table. Bedrock prices by model, by region and by endpoint, and a figure invented for a region nobody checked would be wrong in the direction nobody notices — so everything else raises at preflight and is served with one argument:

from digline.targets import ModelPrice
from digline_bedrock import BedrockTarget, bedrock_pricing

target = BedrockTarget(
    "prompts/answer.md",
    model="amazon.nova-pro-v1:0",
    max_tokens=1024,
    region="us-east-1",
    pricing=bedrock_pricing("us-east-1").override(
        "amazon.nova-pro-v1:0", ModelPrice(input_per_mtok=0.80, output_per_mtok=3.20)
    ),
)

An application inference profile is an ARN and is opaque: it is never in the list, fails preflight, and is served the same way. That is intended, not a gap — a run that cannot say what it cost must not run.

A model you brought in through Custom Model Import, or one behind Provisioned Throughput, has no per-token bill at all: it is billed by model-copy-hour and model-unit-hour. Say so out loud rather than leaving it unpriced:

from digline_bedrock import BedrockTarget, free

target = BedrockTarget(
    "prompts/answer.md",
    model="my-imported-model",
    max_tokens=1024,
    region="eu-west-1",
    pricing=free("my-imported-model"),
)

The details that bite

Your account never leaves the machine. A botocore failure names the assumed role — arn:aws:sts::<account>:assumed-role/… — and digline quotes a target's exception into the reason of every verdict of that case, which lands in a committed run file. So every AWS error is re-raised as BedrockCallFailed with ARNs and account ids removed; the original stays on __cause__, in memory, for a debugger.

Cached tokens. Converse counts cached reads beside inputTokens, not inside them — measured against the API on 2026-08-28, not inferred from the field names: a warm call reported inputTokens=10, cacheReadInputTokens=12002 and totalTokens=12016. So they are added rather than subtracted, which is the opposite of OpenAI's convention. It is one constant in digline_bedrock.client, re-measured by a live test, because getting it wrong is invisible in the direction of good news.

additional_request_fields reaches Converse's additionalModelRequestFields and changes what the model does. It is not part of config_hash — no more than temperature, model or max_tokens are: the fingerprint covers the rules that judge a run, not the system being judged. Two runs differing only in these fields will read as "same configuration as the reference". If you need the difference to show, keep the fields in a file and declare it in Suite.artifacts, and the report carries the diff.

Apache-2.0. Docs: digline/digline.