Merge branch 'main' into fix-nightly-runs-with-haystack-main

integrations/amazon_bedrock/CHANGELOG.md

@@ -1,5 +1,12 @@
 # Changelog
 
+## [integrations/amazon_bedrock-v2.1.1] - 2024-12-18
+
+### 🐛 Bug Fixes
+
+- Fixes to Bedrock Chat Generator for compatibility with the new ChatMessage (#1250)
+
+
 ## [integrations/amazon_bedrock-v2.1.0] - 2024-12-11
 
 ### 🚀 Features

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

@@ -278,12 +278,10 @@ def extract_replies_from_response(self, response_body: Dict[str, Any]) -> List[C
                 # Process each content block separately
                 for content_block in content_blocks:
                     if "text" in content_block:
-                        replies.append(ChatMessage.from_assistant(content=content_block["text"], meta=base_meta.copy()))
+                        replies.append(ChatMessage.from_assistant(content_block["text"], meta=base_meta.copy()))
                     elif "toolUse" in content_block:
                         replies.append(
-                            ChatMessage.from_assistant(
-                                content=json.dumps(content_block["toolUse"]), meta=base_meta.copy()
-                            )
+                            ChatMessage.from_assistant(json.dumps(content_block["toolUse"]), meta=base_meta.copy())
                         )
         return replies
 
@@ -334,9 +332,9 @@ def process_streaming_response(
                         pass
 
                     tool_content = json.dumps(current_tool_use)
-                    replies.append(ChatMessage.from_assistant(content=tool_content, meta=base_meta.copy()))
+                    replies.append(ChatMessage.from_assistant(tool_content, meta=base_meta.copy()))
                 elif current_content:
-                    replies.append(ChatMessage.from_assistant(content=current_content, meta=base_meta.copy()))
+                    replies.append(ChatMessage.from_assistant(current_content, meta=base_meta.copy()))
 
             elif "messageStop" in event:
                 # not 100% correct for multiple messages but no way around it

integrations/amazon_bedrock/tests/test_chat_generator.py

@@ -163,9 +163,9 @@ def test_default_inference_params(self, model_name, chat_messages):
 
         first_reply = replies[0]
         assert isinstance(first_reply, ChatMessage), "First reply is not a ChatMessage instance"
-        assert first_reply.content, "First reply has no content"
+        assert first_reply.text, "First reply has no content"
         assert ChatMessage.is_from(first_reply, ChatRole.ASSISTANT), "First reply is not from the assistant"
-        assert "paris" in first_reply.content.lower(), "First reply does not contain 'paris'"
+        assert "paris" in first_reply.text.lower(), "First reply does not contain 'paris'"
         assert first_reply.meta, "First reply has no metadata"
 
         if first_reply.meta and "usage" in first_reply.meta:
@@ -197,9 +197,9 @@ def streaming_callback(chunk: StreamingChunk):
 
         first_reply = replies[0]
         assert isinstance(first_reply, ChatMessage), "First reply is not a ChatMessage instance"
-        assert first_reply.content, "First reply has no content"
+        assert first_reply.text, "First reply has no content"
         assert ChatMessage.is_from(first_reply, ChatRole.ASSISTANT), "First reply is not from the assistant"
-        assert "paris" in first_reply.content.lower(), "First reply does not contain 'paris'"
+        assert "paris" in first_reply.text.lower(), "First reply does not contain 'paris'"
         assert first_reply.meta, "First reply has no metadata"
 
     @pytest.mark.parametrize("model_name", MODELS_TO_TEST_WITH_TOOLS)
@@ -246,17 +246,17 @@ def test_tools_use(self, model_name):
 
         first_reply = replies[0]
         assert isinstance(first_reply, ChatMessage), "First reply is not a ChatMessage instance"
-        assert first_reply.content, "First reply has no content"
+        assert first_reply.text, "First reply has no content"
         assert ChatMessage.is_from(first_reply, ChatRole.ASSISTANT), "First reply is not from the assistant"
         assert first_reply.meta, "First reply has no metadata"
 
         # Some models return thinking message as first and the second one as the tool call
         if len(replies) > 1:
             second_reply = replies[1]
             assert isinstance(second_reply, ChatMessage), "Second reply is not a ChatMessage instance"
-            assert second_reply.content, "Second reply has no content"
+            assert second_reply.text, "Second reply has no content"
             assert ChatMessage.is_from(second_reply, ChatRole.ASSISTANT), "Second reply is not from the assistant"
-            tool_call = json.loads(second_reply.content)
+            tool_call = json.loads(second_reply.text)
             assert "toolUseId" in tool_call, "Tool call does not contain 'toolUseId' key"
             assert tool_call["name"] == "top_song", f"Tool call {tool_call} does not contain the correct 'name' value"
             assert "input" in tool_call, f"Tool call {tool_call} does not contain 'input' key"
@@ -266,7 +266,7 @@ def test_tools_use(self, model_name):
         else:
             # case where the model returns the tool call as the first message
             # double check that the tool call is correct
-            tool_call = json.loads(first_reply.content)
+            tool_call = json.loads(first_reply.text)
             assert "toolUseId" in tool_call, "Tool call does not contain 'toolUseId' key"
             assert tool_call["name"] == "top_song", f"Tool call {tool_call} does not contain the correct 'name' value"
             assert "input" in tool_call, f"Tool call {tool_call} does not contain 'input' key"
@@ -318,17 +318,17 @@ def test_tools_use_with_streaming(self, model_name):
 
         first_reply = replies[0]
         assert isinstance(first_reply, ChatMessage), "First reply is not a ChatMessage instance"
-        assert first_reply.content, "First reply has no content"
+        assert first_reply.text, "First reply has no content"
         assert ChatMessage.is_from(first_reply, ChatRole.ASSISTANT), "First reply is not from the assistant"
         assert first_reply.meta, "First reply has no metadata"
 
         # Some models return thinking message as first and the second one as the tool call
         if len(replies) > 1:
             second_reply = replies[1]
             assert isinstance(second_reply, ChatMessage), "Second reply is not a ChatMessage instance"
-            assert second_reply.content, "Second reply has no content"
+            assert second_reply.text, "Second reply has no content"
             assert ChatMessage.is_from(second_reply, ChatRole.ASSISTANT), "Second reply is not from the assistant"
-            tool_call = json.loads(second_reply.content)
+            tool_call = json.loads(second_reply.text)
             assert "toolUseId" in tool_call, "Tool call does not contain 'toolUseId' key"
             assert tool_call["name"] == "top_song", f"Tool call {tool_call} does not contain the correct 'name' value"
             assert "input" in tool_call, f"Tool call {tool_call} does not contain 'input' key"
@@ -338,7 +338,7 @@ def test_tools_use_with_streaming(self, model_name):
         else:
             # case where the model returns the tool call as the first message
             # double check that the tool call is correct
-            tool_call = json.loads(first_reply.content)
+            tool_call = json.loads(first_reply.text)
             assert "toolUseId" in tool_call, "Tool call does not contain 'toolUseId' key"
             assert tool_call["name"] == "top_song", f"Tool call {tool_call} does not contain the correct 'name' value"
             assert "input" in tool_call, f"Tool call {tool_call} does not contain 'input' key"
@@ -361,7 +361,7 @@ def test_extract_replies_from_response(self, mock_boto3_session):
 
         replies = generator.extract_replies_from_response(text_response)
         assert len(replies) == 1
-        assert replies[0].content == "This is a test response"
+        assert replies[0].text == "This is a test response"
         assert replies[0].role == ChatRole.ASSISTANT
         assert replies[0].meta["model"] == "anthropic.claude-3-5-sonnet-20240620-v1:0"
         assert replies[0].meta["finish_reason"] == "complete"
@@ -381,7 +381,7 @@ def test_extract_replies_from_response(self, mock_boto3_session):
 
         replies = generator.extract_replies_from_response(tool_response)
         assert len(replies) == 1
-        tool_content = json.loads(replies[0].content)
+        tool_content = json.loads(replies[0].text)
         assert tool_content["toolUseId"] == "123"
         assert tool_content["name"] == "test_tool"
         assert tool_content["input"] == {"key": "value"}
@@ -405,8 +405,8 @@ def test_extract_replies_from_response(self, mock_boto3_session):
 
         replies = generator.extract_replies_from_response(mixed_response)
         assert len(replies) == 2
-        assert replies[0].content == "Let me help you with that. I'll use the search tool to find the answer."
-        tool_content = json.loads(replies[1].content)
+        assert replies[0].text == "Let me help you with that. I'll use the search tool to find the answer."
+        tool_content = json.loads(replies[1].text)
         assert tool_content["toolUseId"] == "456"
         assert tool_content["name"] == "search_tool"
         assert tool_content["input"] == {"query": "test"}
@@ -446,13 +446,13 @@ def test_callback(chunk: StreamingChunk):
         # Verify final replies
         assert len(replies) == 2
         # Check text reply
-        assert replies[0].content == "Let me help you."
+        assert replies[0].text == "Let me help you."
         assert replies[0].meta["model"] == "anthropic.claude-3-5-sonnet-20240620-v1:0"
         assert replies[0].meta["finish_reason"] == "complete"
         assert replies[0].meta["usage"] == {"prompt_tokens": 10, "completion_tokens": 20, "total_tokens": 30}
 
         # Check tool use reply
-        tool_content = json.loads(replies[1].content)
+        tool_content = json.loads(replies[1].text)
         assert tool_content["toolUseId"] == "123"
         assert tool_content["name"] == "search_tool"
         assert tool_content["input"] == {"query": "test"}

integrations/anthropic/CHANGELOG.md

@@ -1,6 +1,10 @@
 # Changelog
 
-## [unreleased]
+## [integrations/anthropic-v1.2.1] - 2024-12-18
+
+### 🐛 Bug Fixes
+
+- Make Anthropic compatible with new `ChatMessage`; fix prompt caching tests (#1252)
 
 ### ⚙️ CI
 
@@ -9,10 +13,13 @@
 ### 🧹 Chores
 
 - Update ruff linting scripts and settings (#1105)
+- Fix linting/isort (#1215)
 
 ### 🌀 Miscellaneous
 
 - Add AnthropicVertexChatGenerator component (#1192)
+- Docs: add AnthropicVertexChatGenerator to pydoc (#1221)
+- Chore: use `text` instead of `content` for `ChatMessage` in Cohere and Anthropic (#1237)
 
 ## [integrations/anthropic-v1.1.0] - 2024-09-20
 

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

@@ -1,4 +1,3 @@
-import dataclasses
 import json
 from typing import Any, Callable, ClassVar, Dict, List, Optional, Union
 
@@ -275,8 +274,16 @@ def _convert_to_anthropic_format(self, messages: List[ChatMessage]) -> List[Dict
         """
         anthropic_formatted_messages = []
         for m in messages:
-            message_dict = dataclasses.asdict(m)
-            formatted_message = {k: v for k, v in message_dict.items() if k in {"role", "content"} and v}
+            message_dict = m.to_dict()
+            formatted_message = {}
+
+            # legacy format
+            if "role" in message_dict and "content" in message_dict:
+                formatted_message = {k: v for k, v in message_dict.items() if k in {"role", "content"} and v}
+            # new format
+            elif "_role" in message_dict and "_content" in message_dict:
+                formatted_message = {"role": m.role.value, "content": m.text}
+
             if m.is_from(ChatRole.SYSTEM):
                 # system messages are treated differently and MUST be in the format expected by the Anthropic API
                 # remove role and content from the message dict, add type and text

integrations/anthropic/tests/test_chat_generator.py

@@ -428,5 +428,5 @@ def test_prompt_caching(self, cache_enabled):
                 or token_usage.get("cache_read_input_tokens") > 1024
             )
         else:
-            assert "cache_creation_input_tokens" not in token_usage
-            assert "cache_read_input_tokens" not in token_usage
+            assert token_usage["cache_creation_input_tokens"] == 0
+            assert token_usage["cache_read_input_tokens"] == 0