You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Bedrock providers ignore AWS_ENDPOINT_URL_BEDROCK_RUNTIME env var #415
ChatBedrockAnthropic/ChatAWSBedrock (chatlas/_provider_bedrock.py) and the Converse-based provider (chatlas/_provider_bedrock_converse.py) both compute base_url from the AWS region whenever no explicit base_url is passed:
# _provider_bedrock.py / _provider_bedrock_converse.pyresolved_base_url=base_urlorbedrock_base_url("converse", aws_region) # or "messages"/"responses"
This bypasses botocore's native support for AWS_ENDPOINT_URL_BEDROCK_RUNTIME (and the generic AWS_ENDPOINT_URL), which is the standard way to point AWS SDKs at a custom Bedrock runtime endpoint (e.g. VPC endpoints, proxies, local testing, FIPS endpoints, or region-specific overrides not covered by the built-in formula).
ellmer just added this: chat_aws_bedrock() now defaults base_url from AWS_ENDPOINT_URL_BEDROCK_RUNTIME (tidyverse/ellmer#1104). The corresponding Anthropic-direct case, ANTHROPIC_BASE_URL, is already handled for free in chatlas since the anthropic Python SDK reads that env var internally -- no action needed there.
Expected fix
In _provider_bedrock.py and _provider_bedrock_converse.py, when base_url is not explicitly passed, check os.environ.get("AWS_ENDPOINT_URL_BEDROCK_RUNTIME") before falling back to the computed bedrock_base_url(...) formula.
Activity
fix(bedrock): honor generic AWS_ENDPOINT_URL with botocore precedence