Print prompt caching data in example

integrations/anthropic/example/prompt_caching.py

@@ -53,11 +53,13 @@
 # Answer the questions using prompt caching (i.e. the entire document is cached, we run the question against it)
 for question in questions:
     print("Question: " + question)
-    qa_pipeline.run(
+    result = qa_pipeline.run(
         data={
             "llm": {"messages": [final_prompt_msg,
                                  ChatMessage.from_user("Given these documents, answer the question:" + question)]},
         }
     )
+
+    print("\n\nChecking cache usage:", result["llm"]["replies"][0].meta.get("usage"))
     print("\n")
 

integrations/anthropic/src/haystack_integrations/components/generators/anthropic/chat/chat_generator.py

@@ -184,7 +184,7 @@ def run(self, messages: List[ChatMessage], generation_kwargs: Optional[Dict[str,
         anthropic_formatted_messages = self._convert_to_anthropic_format(messages)
 
         # system message provided by the user overrides the system message from the self.generation_kwargs
-        system = messages[0].content if messages and messages[0].is_from(ChatRole.SYSTEM) else None
+        system = [anthropic_formatted_messages[0]] if messages and messages[0].is_from(ChatRole.SYSTEM) else None
         if system:
             anthropic_formatted_messages = anthropic_formatted_messages[1:]
 
@@ -262,6 +262,11 @@ def _convert_to_anthropic_format(self, messages: List[ChatMessage]) -> List[Dict
         for m in messages:
             message_dict = dataclasses.asdict(m)
             filtered_message = {k: v for k, v in message_dict.items() if k in {"role", "content"} and v}
+            if m.is_from(ChatRole.SYSTEM):
+                # system messages need to be in the format expected by the Anthropic API
+                filtered_message.pop("role")
+                filtered_message["type"] = "text"
+                filtered_message["text"] = filtered_message.pop("content")
             filtered_message.update(m.meta or {})
             anthropic_formatted_messages.append(filtered_message)
         return anthropic_formatted_messages