[fix] serval bugs in bedrock LLM

metagpt/provider/bedrock/base_provider.py

@@ -25,4 +25,4 @@ class BaseBedrockProvider(ABC):
 
     def messages_to_prompt(self, messages: list[dict]) -> str:
         """[{"role": "user", "content": msg}] to user: <msg> etc."""
-        return "\n".join([f"{i['role']}: {i['content']}" for i in messages])
+        return "\n".join([f"{msg['role']}: {msg['content']}" for msg in messages])

metagpt/provider/bedrock/bedrock_provider.py

@@ -1,5 +1,5 @@
 import json
-from typing import Literal
+from typing import Literal, Tuple
 
 from metagpt.provider.bedrock.base_provider import BaseBedrockProvider
 from metagpt.provider.bedrock.utils import (
@@ -21,8 +21,19 @@ class MistralProvider(BaseBedrockProvider):
 class AnthropicProvider(BaseBedrockProvider):
     # See https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-anthropic-claude-messages.html
 
-    def get_request_body(self, messages: list[dict], generate_kwargs, *args, **kwargs):
-        body = json.dumps({"messages": messages, "anthropic_version": "bedrock-2023-05-31", **generate_kwargs})
+    def _split_system_user_messages(self, messages: list[dict]) -> Tuple[str, list[dict]]:
+        system_messages = []
+        user_messages = []
+        for message in messages:
+            if message["role"] == "system":
+                system_messages.append(message)
+            else:
+                user_messages.append(message)
+        return self.messages_to_prompt(system_messages), user_messages
+
+    def get_request_body(self, messages: list[dict], generate_kwargs, *args, **kwargs) -> str:
+        system_message, user_messages = self._split_system_user_messages(messages)
+        body = json.dumps({"messages": user_messages, "anthropic_version": "bedrock-2023-05-31", "system": system_message, **generate_kwargs})
         return body
 
     def _get_completion_from_dict(self, rsp_dict: dict) -> str:

metagpt/provider/bedrock_api.py

@@ -131,10 +131,8 @@ class BedrockLLM(BaseLLM):
         headers = response.get("ResponseMetadata", {}).get("HTTPHeaders", {})
         prompt_tokens = int(headers.get("x-amzn-bedrock-input-token-count", 0))
         completion_tokens = int(headers.get("x-amzn-bedrock-output-token-count", 0))
-        usage = (
-            {
+        usage = {
                 "prompt_tokens": prompt_tokens,
                 "completion_tokens": completion_tokens,
-            },
-        )
+            }
         return usage