AWS Bedrock
Claude on AWS Bedrock via the Converse API, selected with LLM_PROVIDER=bedrock, used to extract facts from conversations.
Enable it
bash
LLM_PROVIDER=bedrock
BEDROCK_LLM_MODEL=global.anthropic.claude-haiku-4-5-20251001-v1:0
BEDROCK_REGION=ap-south-1
# Credentials come from the ambient AWS identity (ECS task role or AWS_PROFILE) — no API key.Configuration
| Variable | Default | Description |
|---|---|---|
LLM_PROVIDER | openai | Set to bedrock. |
BEDROCK_LLM_MODEL | global.anthropic.claude-haiku-4-5-20251001-v1:0 | Inference-profile id for the Converse model. |
BEDROCK_REGION | ap-south-1 | AWS region for the Bedrock runtime client. |
BEDROCK_CHAT_MODEL | (none) | Optional stronger model for user-facing chat answers; falls back to BEDROCK_LLM_MODEL. |
BEDROCK_REFLECTION_MODEL | (none) | Optional model for reflection synthesis; falls back to BEDROCK_LLM_MODEL. |
Notes
- No API key: credentials come from the ambient AWS identity (ECS task role in staging/prod,
AWS_PROFILElocally). - Current-generation Claude requires an inference-profile id (e.g. a
global.orapac.prefix); a bare model id is rejected with a ValidationException. - Converse has no JSON mode, so JSON requests use an assistant prefill (forcing the response to continue from
{); models that reject prefill retry once without it.