diff --git a/deploy/docker-compose.prod.yml b/deploy/docker-compose.prod.yml index 3ee5217..cd3b90c 100644 --- a/deploy/docker-compose.prod.yml +++ b/deploy/docker-compose.prod.yml @@ -246,8 +246,10 @@ services: LOCAL_CPU_PROVIDER_URL: http://docker.cin.su:18082/v1 LOCAL_CPU_PROVIDER_API_KEY_FILE: ${LOCAL_CPU_PROVIDER_API_KEY_FILE:-} LOCAL_CPU_PROVIDER_CREDENTIAL_REF: ${LOCAL_CPU_PROVIDER_CREDENTIAL_REF:-credential://domovoy/local-cpu/provider-api-key/v1} - LOCAL_MEDIUM_PROVIDER_URL: ${LOCAL_MEDIUM_PROVIDER_URL:-http://llm-local-api:8080/v1} - LOCAL_MEDIUM_PROVIDER_API_KEY_FILE: ${LOCAL_MEDIUM_PROVIDER_API_KEY_FILE:-} + # The medium chat route is served by the authenticated local gateway. + # llm-local-api is the embeddings service and must not be used for chat. + LOCAL_MEDIUM_PROVIDER_URL: ${LOCAL_MEDIUM_PROVIDER_URL:-http://llm-local-gateway:8080/v1} + LOCAL_MEDIUM_PROVIDER_API_KEY_FILE: ${LOCAL_MEDIUM_PROVIDER_API_KEY_FILE:-/run/secrets/llm_local_gateway_api_key} LOCAL_MEDIUM_PROVIDER_CREDENTIAL_REF: ${LOCAL_MEDIUM_PROVIDER_CREDENTIAL_REF:-credential://domovoy/local-medium/provider-api-key/v1} LOCAL_GPU_PROVIDER_TIMEOUT: ${LOCAL_GPU_PROVIDER_TIMEOUT:-60s} LOCAL_LIFEBOAT_PROVIDER_URL: http://cognitive-lifeboat:8086/v1