Fix message history limiter for tool call (microsoft#3178)

* fix: message history limiter to support tool calls

* add: pytest and docs for message history limiter for tool calls

* Added keep_first_message for HistoryLimiter transform

* Update to inbetween to between

* Updated keep_first_message to non-optional, logic for history limiter

* Update transforms.py

* Update test_transforms to match utils introduction, add keep_first_message testing

* Update test_transforms.py for pre-commit checks

---------

Co-authored-by: Mark Sze <[email protected]>
Co-authored-by: Chi Wang <[email protected]>

autogen/agentchat/contrib/capabilities/transforms.py

@@ -53,13 +53,16 @@ class MessageHistoryLimiter:
     It trims the conversation history by removing older messages, retaining only the most recent messages.
     """
 
-    def __init__(self, max_messages: Optional[int] = None):
+    def __init__(self, max_messages: Optional[int] = None, keep_first_message: bool = False):
         """
         Args:
             max_messages Optional[int]: Maximum number of messages to keep in the context. Must be greater than 0 if not None.
+            keep_first_message bool: Whether to keep the original first message in the conversation history.
+                Defaults to False.
         """
         self._validate_max_messages(max_messages)
         self._max_messages = max_messages
+        self._keep_first_message = keep_first_message
 
     def apply_transform(self, messages: List[Dict]) -> List[Dict]:
         """Truncates the conversation history to the specified maximum number of messages.
@@ -75,10 +78,31 @@ def apply_transform(self, messages: List[Dict]) -> List[Dict]:
             List[Dict]: A new list containing the most recent messages up to the specified maximum.
         """
 
-        if self._max_messages is None:
+        if self._max_messages is None or len(messages) <= self._max_messages:
             return messages
 
-        return messages[-self._max_messages :]
+        truncated_messages = []
+        remaining_count = self._max_messages
+
+        # Start with the first message if we need to keep it
+        if self._keep_first_message:
+            truncated_messages = [messages[0]]
+            remaining_count -= 1
+
+        # Loop through messages in reverse
+        for i in range(len(messages) - 1, 0, -1):
+            if remaining_count > 1:
+                truncated_messages.insert(1 if self._keep_first_message else 0, messages[i])
+            if remaining_count == 1:
+                # If there's only 1 slot left and it's a 'tools' message, ignore it.
+                if messages[i].get("role") != "tool":
+                    truncated_messages.insert(1, messages[i])
+
+            remaining_count -= 1
+            if remaining_count == 0:
+                break
+
+        return truncated_messages
 
     def get_logs(self, pre_transform_messages: List[Dict], post_transform_messages: List[Dict]) -> Tuple[str, bool]:
         pre_transform_messages_len = len(pre_transform_messages)

test/agentchat/contrib/capabilities/test_transforms.py

@@ -9,8 +9,8 @@
     MessageHistoryLimiter,
     MessageTokenLimiter,
     TextMessageCompressor,
-    _count_tokens,
 )
+from autogen.agentchat.contrib.capabilities.transforms_util import count_text_tokens
 
 
 class _MockTextCompressor:
@@ -40,6 +40,26 @@ def get_no_content_messages() -> List[Dict]:
     return [{"role": "user", "function_call": "example"}, {"role": "assistant", "content": None}]
 
 
+def get_tool_messages() -> List[Dict]:
+    return [
+        {"role": "user", "content": "hello"},
+        {"role": "tool_calls", "content": "calling_tool"},
+        {"role": "tool", "content": "tool_response"},
+        {"role": "user", "content": "how are you"},
+        {"role": "assistant", "content": [{"type": "text", "text": "are you doing?"}]},
+    ]
+
+
+def get_tool_messages_kept() -> List[Dict]:
+    return [
+        {"role": "user", "content": "hello"},
+        {"role": "tool_calls", "content": "calling_tool"},
+        {"role": "tool", "content": "tool_response"},
+        {"role": "tool_calls", "content": "calling_tool"},
+        {"role": "tool", "content": "tool_response"},
+    ]
+
+
 def get_text_compressors() -> List[TextCompressor]:
     compressors: List[TextCompressor] = [_MockTextCompressor()]
     try:
@@ -57,6 +77,11 @@ def message_history_limiter() -> MessageHistoryLimiter:
     return MessageHistoryLimiter(max_messages=3)
 
 
+@pytest.fixture
+def message_history_limiter_keep_first() -> MessageHistoryLimiter:
+    return MessageHistoryLimiter(max_messages=3, keep_first_message=True)
+
+
 @pytest.fixture
 def message_token_limiter() -> MessageTokenLimiter:
     return MessageTokenLimiter(max_tokens_per_message=3)
@@ -96,19 +121,52 @@ def _filter_dict_test(
 
 @pytest.mark.parametrize(
     "messages, expected_messages_len",
-    [(get_long_messages(), 3), (get_short_messages(), 3), (get_no_content_messages(), 2)],
+    [
+        (get_long_messages(), 3),
+        (get_short_messages(), 3),
+        (get_no_content_messages(), 2),
+        (get_tool_messages(), 2),
+        (get_tool_messages_kept(), 2),
+    ],
 )
 def test_message_history_limiter_apply_transform(message_history_limiter, messages, expected_messages_len):
     transformed_messages = message_history_limiter.apply_transform(messages)
     assert len(transformed_messages) == expected_messages_len
 
+    if messages == get_tool_messages_kept():
+        assert transformed_messages[0]["role"] == "tool_calls"
+        assert transformed_messages[1]["role"] == "tool"
+
+
+@pytest.mark.parametrize(
+    "messages, expected_messages_len",
+    [
+        (get_long_messages(), 3),
+        (get_short_messages(), 3),
+        (get_no_content_messages(), 2),
+        (get_tool_messages(), 3),
+        (get_tool_messages_kept(), 3),
+    ],
+)
+def test_message_history_limiter_apply_transform_keep_first(
+    message_history_limiter_keep_first, messages, expected_messages_len
+):
+    transformed_messages = message_history_limiter_keep_first.apply_transform(messages)
+    assert len(transformed_messages) == expected_messages_len
+
+    if messages == get_tool_messages_kept():
+        assert transformed_messages[1]["role"] == "tool_calls"
+        assert transformed_messages[2]["role"] == "tool"
+
 
 @pytest.mark.parametrize(
     "messages, expected_logs, expected_effect",
     [
         (get_long_messages(), "Removed 2 messages. Number of messages reduced from 5 to 3.", True),
         (get_short_messages(), "No messages were removed.", False),
         (get_no_content_messages(), "No messages were removed.", False),
+        (get_tool_messages(), "Removed 3 messages. Number of messages reduced from 5 to 2.", True),
+        (get_tool_messages_kept(), "Removed 3 messages. Number of messages reduced from 5 to 2.", True),
     ],
 )
 def test_message_history_limiter_get_logs(message_history_limiter, messages, expected_logs, expected_effect):
@@ -131,7 +189,8 @@ def test_message_token_limiter_apply_transform(
 ):
     transformed_messages = message_token_limiter.apply_transform(copy.deepcopy(messages))
     assert (
-        sum(_count_tokens(msg["content"]) for msg in transformed_messages if "content" in msg) == expected_token_count
+        sum(count_text_tokens(msg["content"]) for msg in transformed_messages if "content" in msg)
+        == expected_token_count
     )
     assert len(transformed_messages) == expected_messages_len
 
@@ -167,7 +226,8 @@ def test_message_token_limiter_with_threshold_apply_transform(
 ):
     transformed_messages = message_token_limiter_with_threshold.apply_transform(messages)
     assert (
-        sum(_count_tokens(msg["content"]) for msg in transformed_messages if "content" in msg) == expected_token_count
+        sum(count_text_tokens(msg["content"]) for msg in transformed_messages if "content" in msg)
+        == expected_token_count
     )
     assert len(transformed_messages) == expected_messages_len
 
@@ -240,56 +300,31 @@ def test_text_compression_with_filter(messages, text_compressor):
         assert _filter_dict_test(post_transform, pre_transform, ["user"], exclude_filter=False)
 
 
-@pytest.mark.parametrize("text_compressor", get_text_compressors())
-def test_text_compression_cache(text_compressor):
-    messages = get_long_messages()
-    mock_compressed_content = (1, {"content": "mock"})
-
-    with patch(
-        "autogen.agentchat.contrib.capabilities.transforms.TextMessageCompressor._cache_get",
-        MagicMock(return_value=(1, {"content": "mock"})),
-    ) as mocked_get, patch(
-        "autogen.agentchat.contrib.capabilities.transforms.TextMessageCompressor._cache_set", MagicMock()
-    ) as mocked_set:
-        compressor = TextMessageCompressor(text_compressor=text_compressor)
-
-        compressor.apply_transform(messages)
-        compressor.apply_transform(messages)
-
-        assert mocked_get.call_count == len(messages)
-        assert mocked_set.call_count == len(messages)
-
-    # We already populated the cache with the mock content
-    # We need to test if we retrieve the correct content
-    compressor = TextMessageCompressor(text_compressor=text_compressor)
-    compressed_messages = compressor.apply_transform(messages)
-
-    for message in compressed_messages:
-        assert message["content"] == mock_compressed_content[1]
-
-
 if __name__ == "__main__":
     long_messages = get_long_messages()
     short_messages = get_short_messages()
     no_content_messages = get_no_content_messages()
+    tool_messages = get_tool_messages()
     msg_history_limiter = MessageHistoryLimiter(max_messages=3)
+    msg_history_limiter_keep_first = MessageHistoryLimiter(max_messages=3, keep_first=True)
     msg_token_limiter = MessageTokenLimiter(max_tokens_per_message=3)
     msg_token_limiter_with_threshold = MessageTokenLimiter(max_tokens_per_message=1, min_tokens=10)
 
     # Test Parameters
     message_history_limiter_apply_transform_parameters = {
-        "messages": [long_messages, short_messages, no_content_messages],
-        "expected_messages_len": [3, 3, 2],
+        "messages": [long_messages, short_messages, no_content_messages, tool_messages],
+        "expected_messages_len": [3, 3, 2, 4],
     }
 
     message_history_limiter_get_logs_parameters = {
-        "messages": [long_messages, short_messages, no_content_messages],
+        "messages": [long_messages, short_messages, no_content_messages, tool_messages],
         "expected_logs": [
             "Removed 2 messages. Number of messages reduced from 5 to 3.",
             "No messages were removed.",
             "No messages were removed.",
+            "Removed 1 messages. Number of messages reduced from 5 to 4.",
         ],
-        "expected_effect": [True, False, False],
+        "expected_effect": [True, False, False, True],
     }
 
     message_token_limiter_apply_transform_parameters = {
@@ -322,6 +357,14 @@ def test_text_compression_cache(text_compressor):
     ):
         test_message_history_limiter_apply_transform(msg_history_limiter, messages, expected_messages_len)
 
+    for messages, expected_messages_len in zip(
+        message_history_limiter_apply_transform_parameters["messages"],
+        message_history_limiter_apply_transform_parameters["expected_messages_len"],
+    ):
+        test_message_history_limiter_apply_transform_keep_first(
+            msg_history_limiter_keep_first, messages, expected_messages_len
+        )
+
     for messages, expected_logs, expected_effect in zip(
         message_history_limiter_get_logs_parameters["messages"],
         message_history_limiter_get_logs_parameters["expected_logs"],

website/docs/topics/handling_long_contexts/intro_to_transform_messages.md

@@ -59,7 +59,28 @@ pprint.pprint(processed_messages)
 {'content': 'very very very very very very long string', 'role': 'user'}]
 ```
 
-By applying the `MessageHistoryLimiter`, we can see that we were able to limit the context history to the 3 most recent messages.
+By applying the `MessageHistoryLimiter`, we can see that we were able to limit the context history to the 3 most recent messages. However, if the splitting point is between a "tool_calls" and "tool" pair, the complete pair will be included to obey the OpenAI API call constraints.
+
+```python
+max_msg_transfrom = transforms.MessageHistoryLimiter(max_messages=3)
+
+messages = [
+    {"role": "user", "content": "hello"},
+    {"role": "tool_calls", "content": "calling_tool"},
+    {"role": "tool", "content": "tool_response"},
+    {"role": "user", "content": "how are you"},
+    {"role": "assistant", "content": [{"type": "text", "text": "are you doing?"}]},
+]
+
+processed_messages = max_msg_transfrom.apply_transform(copy.deepcopy(messages))
+pprint.pprint(processed_messages)
+```
+```console
+[{'content': 'calling_tool', 'role': 'tool_calls'},
+{'content': 'tool_response', 'role': 'tool'},
+{'content': 'how are you', 'role': 'user'},
+{'content': [{'text': 'are you doing?', 'type': 'text'}], 'role': 'assistant'}]
+```
 
 #### Example 2: Limiting the Number of Tokens