Skip to content

bedrock.md

Latest commit

 

History

History
122 lines (92 loc) · 3.61 KB

bedrock.md

File metadata and controls

122 lines (92 loc) · 3.61 KB
Raw

Integration Guide: AWS Bedrock Converse

Use Skillware skills with models hosted on Amazon Bedrock via the Converse API and native tool-use schema.

Install

pip install "skillware[bedrock]"
pip install "skillware[compliance_tos_evaluator]"   # example skill extra

Installs boto3. Configure AWS credentials (IAM role on EC2/ECS/Lambda, or AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY locally) and set AWS_REGION.

Adapter

from skillware.core.loader import SkillLoader

bundle = SkillLoader.load_skill("compliance/tos_evaluator")
tool = SkillLoader.to_bedrock_tool(bundle)
# tool → {"toolSpec": {"name", "description", "inputSchema": {"json": ...}}}

Pass tool inside toolConfig.tools when calling bedrock-runtime converse(). Tool names use the same sanitization as OpenAI adapters (finance/wallet_screening → finance_wallet_screening).

Claude, Llama, and other models on Bedrock share this Converse tool shape — you do not need a separate adapter per Bedrock model family.

Minimal loop

import os

import boto3

from skillware.core.env import load_env_file
from skillware.core.loader import SkillLoader

load_env_file()

bundle = SkillLoader.load_skill("compliance/tos_evaluator")
skill = bundle["class"]()
tool = SkillLoader.to_bedrock_tool(bundle)
tool_name = tool["toolSpec"]["name"]

client = boto3.client("bedrock-runtime", region_name=os.environ["AWS_REGION"])
model_id = os.environ["BEDROCK_MODEL_ID"]

messages = [
    {
        "role": "user",
        "content": [{"text": "Check whether https://example.com/docs allows crawling."}],
    }
]

response = client.converse(
    modelId=model_id,
    system=[{"text": bundle["instructions"]}],
    messages=messages,
    toolConfig={"tools": [tool], "toolChoice": {"auto": {}}},
)

while True:
    output = response["output"]["message"]
    messages.append(output)
    tool_uses = [block for block in output["content"] if "toolUse" in block]
    if not tool_uses:
        break

    tool_results = []
    for block in tool_uses:
        use = block["toolUse"]
        if use["name"] != tool_name:
            raise RuntimeError(f"Unexpected tool: {use['name']}")
        result = skill.execute(dict(use["input"]))
        tool_results.append(
            {
                "toolResult": {
                    "toolUseId": use["toolUseId"],
                    "content": [{"json": result}],
                }
            }
        )

    messages.append({"role": "user", "content": tool_results})
    response = client.converse(
        modelId=model_id,
        system=[{"text": bundle["instructions"]}],
        messages=messages,
        toolConfig={"tools": [tool], "toolChoice": {"auto": {}}},
    )

for block in response["output"]["message"]["content"]:
    if "text" in block:
        print(block["text"])

Runnable copy: examples/bedrock_tos_evaluator.py.

Environment

Variable Purpose
AWS_REGION Bedrock region (for example us-east-1)
BEDROCK_MODEL_ID Model ID (for example anthropic.claude-3-5-sonnet-20241022-v2:0)
Skill env_vars Separate from AWS — see the skill catalog page

Multi-skill hosts

from skillware import SkillContext

ctx = SkillContext(categories=["compliance"], mode="brief")
tools = ctx.tools("bedrock")

Related