Merge pull request #1 from Constantinople-AI/GGGG-29-aws-retry

GGGG-29 Configurable AWS retries via environment variable

Author: ofawx

src/api/models/bedrock.py

@@ -38,11 +38,15 @@
     Usage,
     UserMessage,
 )
-from api.setting import AWS_REGION, CUSTOM_MODEL_LIST, DEBUG, DEFAULT_MODEL, ENABLE_CROSS_REGION_INFERENCE
+from api.setting import AWS_REGION, CUSTOM_MODEL_LIST, DEBUG, DEFAULT_MODEL, ENABLE_CROSS_REGION_INFERENCE, MAX_RETRIES_AWS
 
 logger = logging.getLogger(__name__)
 
-config = Config(connect_timeout=60, read_timeout=120, retries={"max_attempts": 1})
+config = Config(
+    connect_timeout=60,
+    read_timeout=120,
+    retries={"mode": "standard", "max_attempts": MAX_RETRIES_AWS}
+)
 
 bedrock_runtime = boto3.client(
     service_name="bedrock-runtime",

src/api/setting.py

@@ -18,3 +18,4 @@
 CUSTOM_MODEL_CSV = os.environ.get("CUSTOM_MODEL_LIST", "")
 CUSTOM_MODEL_LIST = [m.strip() for m in CUSTOM_MODEL_CSV.split(",") if m.strip()]
 ENABLE_CROSS_REGION_INFERENCE = os.environ.get("ENABLE_CROSS_REGION_INFERENCE", "true").lower() != "false"
+MAX_RETRIES_AWS = int(os.environ.get("MAX_RETRIES_AWS", "3"))