From 932a3e7a472d79b1cb704489e06a73dd534e31cf Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E7=8E=8B=E7=A5=A5=E5=AE=87?= <625024108@qq.com>
Date: Wed, 30 Jul 2025 15:13:14 +0800
Subject: [PATCH 1/5] fix bugs truncate_memory_with_tokenizer

---
 swarms/structs/conversation.py | 119 +++++++++++++++++++++++++++++----
 1 file changed, 105 insertions(+), 14 deletions(-)

diff --git a/swarms/structs/conversation.py b/swarms/structs/conversation.py
index 7c8d3109..3e47cae1 100644
--- a/swarms/structs/conversation.py
+++ b/swarms/structs/conversation.py
@@ -184,6 +184,7 @@ class Conversation:
         system_prompt: Optional[str] = None,
         time_enabled: bool = False,
         autosave: bool = False,  # Changed default to False
+        save_enabled: bool = False,  # New parameter to control if saving is enabled
         save_filepath: str = None,
         load_filepath: str = None,  # New parameter to specify which file to load from
         context_length: int = 8192,
@@ -222,6 +223,7 @@ class Conversation:
         self.system_prompt = system_prompt
         self.time_enabled = time_enabled
         self.autosave = autosave
+        self.save_enabled = save_enabled
         self.conversations_dir = conversations_dir
         self.tokenizer_model_name = tokenizer_model_name
         self.message_id_on = message_id_on
@@ -1019,6 +1021,13 @@ class Conversation:
                 )
                 return
 
+            # Don't save if saving is disabled (你的PR代码)
+            if not self.save_enabled:
+                logger.warning(
+                    "An attempt to save the conversation failed: save_enabled is False."
+                    "Please set save_enabled=True when creating a Conversation object to enable saving."
+                )
+                return
             # Get the full data including metadata and conversation history
             data = self.get_init_params()
 
@@ -1267,39 +1276,121 @@ class Conversation:
 
     def truncate_memory_with_tokenizer(self):
         """
-        Truncates the conversation history based on the total number of tokens using a tokenizer.
-
+        Truncate conversation history based on the total token count using tokenizer.
+        
+        This version is more generic, not dependent on a specific LLM model, and can work with any model that provides a counter.
+        Uses count_tokens function to calculate and truncate by message, ensuring the result is still valid content.
+        
         Returns:
             None
         """
+        from swarms.utils.litellm_tokenizer import count_tokens
+        
         total_tokens = 0
         truncated_history = []
-
+        
         for message in self.conversation_history:
             role = message.get("role")
             content = message.get("content")
-            tokens = count_tokens(content, self.tokenizer_model_name)
-            count = tokens  # Assign the token count
-            total_tokens += count
-
-            if total_tokens <= self.context_length:
+            
+            # Convert content to string if it's not already a string
+            if not isinstance(content, str):
+                content = str(content)
+                
+            # Calculate token count for this message
+            token_count = count_tokens(content, self.tokenizer_model_name)
+            
+            # Check if adding this message would exceed the limit
+            if total_tokens + token_count <= self.context_length:
+                # If not exceeding limit, add the full message
                 truncated_history.append(message)
+                total_tokens += token_count
             else:
-                remaining_tokens = self.context_length - (
-                    total_tokens - count
+                # Calculate remaining tokens we can include
+                remaining_tokens = self.context_length - total_tokens
+                
+                # If no token space left, break the loop
+                if remaining_tokens <= 0:
+                    break
+                    
+                # If we have space left, we need to truncate this message
+                # Use binary search to find content length that fits remaining token space
+                truncated_content = self._binary_search_truncate(
+                    content, 
+                    remaining_tokens,
+                    self.tokenizer_model_name
                 )
-                truncated_content = content[
-                    :remaining_tokens
-                ]  # Truncate the content based on the remaining tokens
+                
+                # Create the truncated message
                 truncated_message = {
                     "role": role,
                     "content": truncated_content,
                 }
+                
+                # Add any other fields from the original message
+                for key, value in message.items():
+                    if key not in ["role", "content"]:
+                        truncated_message[key] = value
+                
                 truncated_history.append(truncated_message)
                 break
-
+        
+        # Update conversation history
         self.conversation_history = truncated_history
 
+    def _binary_search_truncate(self, text, target_tokens, model_name):
+        """
+        Use binary search to find the maximum text substring that fits the target token count.
+        
+        Parameters:
+            text (str): Original text to truncate
+            target_tokens (int): Target token count
+            model_name (str): Model name for token counting
+        
+        Returns:
+            str: Truncated text with token count not exceeding target_tokens
+        """
+        from swarms.utils.litellm_tokenizer import count_tokens
+        
+        # If text is empty or target tokens is 0, return empty string
+        if not text or target_tokens <= 0:
+            return ""
+        
+        # If original text token count is already less than or equal to target, return as is
+        original_tokens = count_tokens(text, model_name)
+        if original_tokens <= target_tokens:
+            return text
+        
+        # Binary search
+        left, right = 0, len(text)
+        best_length = 0
+        best_text = ""
+        
+        while left <= right:
+            mid = (left + right) // 2
+            truncated = text[:mid]
+            tokens = count_tokens(truncated, model_name)
+            
+            if tokens <= target_tokens:
+                # If current truncated text token count is less than or equal to target, try longer text
+                best_length = mid
+                best_text = truncated
+                left = mid + 1
+            else:
+                # Otherwise try shorter text
+                right = mid - 1
+        
+        # Try to truncate at sentence boundaries if possible
+        sentence_delimiters = ['.', '!', '?', '\n']
+        for delimiter in sentence_delimiters:
+            last_pos = best_text.rfind(delimiter)
+            if last_pos > len(best_text) * 0.75:  # Only truncate at sentence boundary if we don't lose too much content
+                truncated_at_sentence = best_text[:last_pos+1]
+                if count_tokens(truncated_at_sentence, model_name) <= target_tokens:
+                    return truncated_at_sentence
+        
+        return best_text
+
     def clear(self):
         """Clear the conversation history."""
         if self.backend_instance:

From 6cdf3d84c8968d08af41796597d4658428826dc2 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E7=8E=8B=E7=A5=A5=E5=AE=87?= <625024108@qq.com>
Date: Thu, 31 Jul 2025 10:07:18 +0800
Subject: [PATCH 2/5] update

---
 examples/utils/misc/conversation_test.py | 103 +++++++++++++++++++----
 1 file changed, 86 insertions(+), 17 deletions(-)

diff --git a/examples/utils/misc/conversation_test.py b/examples/utils/misc/conversation_test.py
index ec8a0534..ae34692b 100644
--- a/examples/utils/misc/conversation_test.py
+++ b/examples/utils/misc/conversation_test.py
@@ -1,22 +1,91 @@
 from swarms.structs.conversation import Conversation
+from dotenv import load_dotenv
 
-# Create a conversation object
-conversation = Conversation(backend="in-memory")
 
-# Add a message to the conversation
-conversation.add(
-    role="user", content="Hello, how are you?", category="input"
-)
+# Load environment variables from .env file
+load_dotenv()
 
-# Add a message to the conversation
-conversation.add(
-    role="assistant",
-    content="I'm good, thank you!",
-    category="output",
-)
-
-print(
-    conversation.export_and_count_categories(
-        tokenizer_model_name="claude-3-5-sonnet-20240620"
+def demonstrate_truncation():
+    # Using a smaller context length to clearly see the truncation effect
+    context_length = 25
+    print(f"Creating a conversation instance with context length {context_length}")
+    
+    # Using Claude model as the tokenizer model
+    conversation = Conversation(
+        context_length=context_length,
+        tokenizer_model_name="claude-3-7-sonnet-20250219"
     )
-)
+    
+    # Adding first message - short message
+    short_message = "Hello, I am a user."
+    print(f"\nAdding short message: '{short_message}'")
+    conversation.add("user", short_message)
+    
+    # Display token count
+    from swarms.utils.litellm_tokenizer import count_tokens
+    tokens = count_tokens(short_message, conversation.tokenizer_model_name)
+    print(f"Short message token count: {tokens}")
+    
+    # Adding second message - long message, should be truncated
+    long_message = "I have a question about artificial intelligence. I want to understand how large language models handle long texts, especially under token constraints. This issue is important because it relates to the model's practicality and effectiveness. I hope to get a detailed answer that helps me understand this complex technical problem."
+    print(f"\nAdding long message:\n'{long_message}'")
+    conversation.add("assistant", long_message)
+    
+    # Display long message token count
+    tokens = count_tokens(long_message, conversation.tokenizer_model_name)
+    print(f"Long message token count: {tokens}")
+    
+    # Display current conversation total token count
+    total_tokens = sum(count_tokens(msg["content"], conversation.tokenizer_model_name) 
+                      for msg in conversation.conversation_history)
+    print(f"Total token count before truncation: {total_tokens}")
+    
+    # Print the complete conversation history before truncation
+    print("\nConversation history before truncation:")
+    for i, msg in enumerate(conversation.conversation_history):
+        print(f"[{i}] {msg['role']}: {msg['content']}")
+        print(f"    Token count: {count_tokens(msg['content'], conversation.tokenizer_model_name)}")
+    
+    # Execute truncation
+    print("\nExecuting truncation...")
+    conversation.truncate_memory_with_tokenizer()
+    
+    # Print conversation history after truncation
+    print("\nConversation history after truncation:")
+    for i, msg in enumerate(conversation.conversation_history):
+        print(f"[{i}] {msg['role']}: {msg['content']}")
+        print(f"    Token count: {count_tokens(msg['content'], conversation.tokenizer_model_name)}")
+    
+    # Display total token count after truncation
+    total_tokens = sum(count_tokens(msg["content"], conversation.tokenizer_model_name) 
+                      for msg in conversation.conversation_history)
+    print(f"\nTotal token count after truncation: {total_tokens}")
+    print(f"Context length limit: {context_length}")
+    
+    # Verify if successfully truncated below the limit
+    if total_tokens <= context_length:
+        print("✅ Success: Total token count is now less than or equal to context length limit")
+    else:
+        print("❌ Failure: Total token count still exceeds context length limit")
+    
+    # Test sentence boundary truncation
+    print("\n\nTesting sentence boundary truncation:")
+    sentence_test = Conversation(context_length=15, tokenizer_model_name="claude-3-opus-20240229")
+    test_text = "This is the first sentence. This is the second very long sentence that contains a lot of content. This is the third sentence."
+    print(f"Original text: '{test_text}'")
+    print(f"Original token count: {count_tokens(test_text, sentence_test.tokenizer_model_name)}")
+    
+    # Using binary search for truncation
+    truncated = sentence_test._binary_search_truncate(test_text, 10, sentence_test.tokenizer_model_name)
+    print(f"Truncated text: '{truncated}'")
+    print(f"Truncated token count: {count_tokens(truncated, sentence_test.tokenizer_model_name)}")
+    
+    # Check if truncated at period
+    if truncated.endswith("."):
+        print("✅ Success: Text was truncated at sentence boundary")
+    else:
+        print("Note: Text was not truncated at sentence boundary")
+
+
+if __name__ == "__main__":
+    demonstrate_truncation()
\ No newline at end of file

From 578452fb6c52989c0ffcb602064ed3c3be23ddf2 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E7=8E=8B=E7=A5=A5=E5=AE=87?= <625024108@qq.com>
Date: Fri, 1 Aug 2025 11:47:21 +0800
Subject: [PATCH 3/5] update

---
 swarms/structs/conversation.py | 17 ++++-------------
 1 file changed, 4 insertions(+), 13 deletions(-)

diff --git a/swarms/structs/conversation.py b/swarms/structs/conversation.py
index 3e47cae1..6e11ebdd 100644
--- a/swarms/structs/conversation.py
+++ b/swarms/structs/conversation.py
@@ -184,7 +184,6 @@ class Conversation:
         system_prompt: Optional[str] = None,
         time_enabled: bool = False,
         autosave: bool = False,  # Changed default to False
-        save_enabled: bool = False,  # New parameter to control if saving is enabled
         save_filepath: str = None,
         load_filepath: str = None,  # New parameter to specify which file to load from
         context_length: int = 8192,
@@ -223,7 +222,6 @@ class Conversation:
         self.system_prompt = system_prompt
         self.time_enabled = time_enabled
         self.autosave = autosave
-        self.save_enabled = save_enabled
         self.conversations_dir = conversations_dir
         self.tokenizer_model_name = tokenizer_model_name
         self.message_id_on = message_id_on
@@ -1021,13 +1019,6 @@ class Conversation:
                 )
                 return
 
-            # Don't save if saving is disabled (你的PR代码)
-            if not self.save_enabled:
-                logger.warning(
-                    "An attempt to save the conversation failed: save_enabled is False."
-                    "Please set save_enabled=True when creating a Conversation object to enable saving."
-                )
-                return
             # Get the full data including metadata and conversation history
             data = self.get_init_params()
 
@@ -1284,7 +1275,7 @@ class Conversation:
         Returns:
             None
         """
-        from swarms.utils.litellm_tokenizer import count_tokens
+        
         
         total_tokens = 0
         truncated_history = []
@@ -1350,7 +1341,7 @@ class Conversation:
         Returns:
             str: Truncated text with token count not exceeding target_tokens
         """
-        from swarms.utils.litellm_tokenizer import count_tokens
+        
         
         # If text is empty or target tokens is 0, return empty string
         if not text or target_tokens <= 0:
@@ -1390,7 +1381,7 @@ class Conversation:
                     return truncated_at_sentence
         
         return best_text
-
+    
     def clear(self):
         """Clear the conversation history."""
         if self.backend_instance:
@@ -1796,4 +1787,4 @@ class Conversation:
 # # # conversation.add("assistant", "I am doing well, thanks.")
 # # # # print(conversation.to_json())
 # # print(type(conversation.to_dict()))
-# # print(conversation.to_yaml())
+# # print(conversation.to_yaml())
\ No newline at end of file

From c9ed0dba2701c78926fff7fab1c5129fbe1d38b9 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E7=8E=8B=E7=A5=A5=E5=AE=87?= <625024108@qq.com>
Date: Sat, 2 Aug 2025 02:10:20 +0800
Subject: [PATCH 4/5] update

---
 examples/utils/misc/conversation_test.py      | 103 +++---------------
 .../utils/misc/conversation_test_truncate.py  |  91 ++++++++++++++++
 2 files changed, 108 insertions(+), 86 deletions(-)
 create mode 100644 examples/utils/misc/conversation_test_truncate.py

diff --git a/examples/utils/misc/conversation_test.py b/examples/utils/misc/conversation_test.py
index ae34692b..a7a6750e 100644
--- a/examples/utils/misc/conversation_test.py
+++ b/examples/utils/misc/conversation_test.py
@@ -1,91 +1,22 @@
 from swarms.structs.conversation import Conversation
-from dotenv import load_dotenv
 
+# Create a conversation object
+conversation = Conversation(backend="in-memory")
 
-# Load environment variables from .env file
-load_dotenv()
-
-def demonstrate_truncation():
-    # Using a smaller context length to clearly see the truncation effect
-    context_length = 25
-    print(f"Creating a conversation instance with context length {context_length}")
-    
-    # Using Claude model as the tokenizer model
-    conversation = Conversation(
-        context_length=context_length,
-        tokenizer_model_name="claude-3-7-sonnet-20250219"
-    )
-    
-    # Adding first message - short message
-    short_message = "Hello, I am a user."
-    print(f"\nAdding short message: '{short_message}'")
-    conversation.add("user", short_message)
-    
-    # Display token count
-    from swarms.utils.litellm_tokenizer import count_tokens
-    tokens = count_tokens(short_message, conversation.tokenizer_model_name)
-    print(f"Short message token count: {tokens}")
-    
-    # Adding second message - long message, should be truncated
-    long_message = "I have a question about artificial intelligence. I want to understand how large language models handle long texts, especially under token constraints. This issue is important because it relates to the model's practicality and effectiveness. I hope to get a detailed answer that helps me understand this complex technical problem."
-    print(f"\nAdding long message:\n'{long_message}'")
-    conversation.add("assistant", long_message)
-    
-    # Display long message token count
-    tokens = count_tokens(long_message, conversation.tokenizer_model_name)
-    print(f"Long message token count: {tokens}")
-    
-    # Display current conversation total token count
-    total_tokens = sum(count_tokens(msg["content"], conversation.tokenizer_model_name) 
-                      for msg in conversation.conversation_history)
-    print(f"Total token count before truncation: {total_tokens}")
-    
-    # Print the complete conversation history before truncation
-    print("\nConversation history before truncation:")
-    for i, msg in enumerate(conversation.conversation_history):
-        print(f"[{i}] {msg['role']}: {msg['content']}")
-        print(f"    Token count: {count_tokens(msg['content'], conversation.tokenizer_model_name)}")
-    
-    # Execute truncation
-    print("\nExecuting truncation...")
-    conversation.truncate_memory_with_tokenizer()
-    
-    # Print conversation history after truncation
-    print("\nConversation history after truncation:")
-    for i, msg in enumerate(conversation.conversation_history):
-        print(f"[{i}] {msg['role']}: {msg['content']}")
-        print(f"    Token count: {count_tokens(msg['content'], conversation.tokenizer_model_name)}")
-    
-    # Display total token count after truncation
-    total_tokens = sum(count_tokens(msg["content"], conversation.tokenizer_model_name) 
-                      for msg in conversation.conversation_history)
-    print(f"\nTotal token count after truncation: {total_tokens}")
-    print(f"Context length limit: {context_length}")
-    
-    # Verify if successfully truncated below the limit
-    if total_tokens <= context_length:
-        print("✅ Success: Total token count is now less than or equal to context length limit")
-    else:
-        print("❌ Failure: Total token count still exceeds context length limit")
-    
-    # Test sentence boundary truncation
-    print("\n\nTesting sentence boundary truncation:")
-    sentence_test = Conversation(context_length=15, tokenizer_model_name="claude-3-opus-20240229")
-    test_text = "This is the first sentence. This is the second very long sentence that contains a lot of content. This is the third sentence."
-    print(f"Original text: '{test_text}'")
-    print(f"Original token count: {count_tokens(test_text, sentence_test.tokenizer_model_name)}")
-    
-    # Using binary search for truncation
-    truncated = sentence_test._binary_search_truncate(test_text, 10, sentence_test.tokenizer_model_name)
-    print(f"Truncated text: '{truncated}'")
-    print(f"Truncated token count: {count_tokens(truncated, sentence_test.tokenizer_model_name)}")
-    
-    # Check if truncated at period
-    if truncated.endswith("."):
-        print("✅ Success: Text was truncated at sentence boundary")
-    else:
-        print("Note: Text was not truncated at sentence boundary")
+# Add a message to the conversation
+conversation.add(
+    role="user", content="Hello, how are you?", category="input"
+)
 
+# Add a message to the conversation
+conversation.add(
+    role="assistant",
+    content="I'm good, thank you!",
+    category="output",
+)
 
-if __name__ == "__main__":
-    demonstrate_truncation()
\ No newline at end of file
+print(
+    conversation.export_and_count_categories(
+        tokenizer_model_name="claude-3-5-sonnet-20240620"
+    )
+)
\ No newline at end of file
diff --git a/examples/utils/misc/conversation_test_truncate.py b/examples/utils/misc/conversation_test_truncate.py
new file mode 100644
index 00000000..ae34692b
--- /dev/null
+++ b/examples/utils/misc/conversation_test_truncate.py
@@ -0,0 +1,91 @@
+from swarms.structs.conversation import Conversation
+from dotenv import load_dotenv
+
+
+# Load environment variables from .env file
+load_dotenv()
+
+def demonstrate_truncation():
+    # Using a smaller context length to clearly see the truncation effect
+    context_length = 25
+    print(f"Creating a conversation instance with context length {context_length}")
+    
+    # Using Claude model as the tokenizer model
+    conversation = Conversation(
+        context_length=context_length,
+        tokenizer_model_name="claude-3-7-sonnet-20250219"
+    )
+    
+    # Adding first message - short message
+    short_message = "Hello, I am a user."
+    print(f"\nAdding short message: '{short_message}'")
+    conversation.add("user", short_message)
+    
+    # Display token count
+    from swarms.utils.litellm_tokenizer import count_tokens
+    tokens = count_tokens(short_message, conversation.tokenizer_model_name)
+    print(f"Short message token count: {tokens}")
+    
+    # Adding second message - long message, should be truncated
+    long_message = "I have a question about artificial intelligence. I want to understand how large language models handle long texts, especially under token constraints. This issue is important because it relates to the model's practicality and effectiveness. I hope to get a detailed answer that helps me understand this complex technical problem."
+    print(f"\nAdding long message:\n'{long_message}'")
+    conversation.add("assistant", long_message)
+    
+    # Display long message token count
+    tokens = count_tokens(long_message, conversation.tokenizer_model_name)
+    print(f"Long message token count: {tokens}")
+    
+    # Display current conversation total token count
+    total_tokens = sum(count_tokens(msg["content"], conversation.tokenizer_model_name) 
+                      for msg in conversation.conversation_history)
+    print(f"Total token count before truncation: {total_tokens}")
+    
+    # Print the complete conversation history before truncation
+    print("\nConversation history before truncation:")
+    for i, msg in enumerate(conversation.conversation_history):
+        print(f"[{i}] {msg['role']}: {msg['content']}")
+        print(f"    Token count: {count_tokens(msg['content'], conversation.tokenizer_model_name)}")
+    
+    # Execute truncation
+    print("\nExecuting truncation...")
+    conversation.truncate_memory_with_tokenizer()
+    
+    # Print conversation history after truncation
+    print("\nConversation history after truncation:")
+    for i, msg in enumerate(conversation.conversation_history):
+        print(f"[{i}] {msg['role']}: {msg['content']}")
+        print(f"    Token count: {count_tokens(msg['content'], conversation.tokenizer_model_name)}")
+    
+    # Display total token count after truncation
+    total_tokens = sum(count_tokens(msg["content"], conversation.tokenizer_model_name) 
+                      for msg in conversation.conversation_history)
+    print(f"\nTotal token count after truncation: {total_tokens}")
+    print(f"Context length limit: {context_length}")
+    
+    # Verify if successfully truncated below the limit
+    if total_tokens <= context_length:
+        print("✅ Success: Total token count is now less than or equal to context length limit")
+    else:
+        print("❌ Failure: Total token count still exceeds context length limit")
+    
+    # Test sentence boundary truncation
+    print("\n\nTesting sentence boundary truncation:")
+    sentence_test = Conversation(context_length=15, tokenizer_model_name="claude-3-opus-20240229")
+    test_text = "This is the first sentence. This is the second very long sentence that contains a lot of content. This is the third sentence."
+    print(f"Original text: '{test_text}'")
+    print(f"Original token count: {count_tokens(test_text, sentence_test.tokenizer_model_name)}")
+    
+    # Using binary search for truncation
+    truncated = sentence_test._binary_search_truncate(test_text, 10, sentence_test.tokenizer_model_name)
+    print(f"Truncated text: '{truncated}'")
+    print(f"Truncated token count: {count_tokens(truncated, sentence_test.tokenizer_model_name)}")
+    
+    # Check if truncated at period
+    if truncated.endswith("."):
+        print("✅ Success: Text was truncated at sentence boundary")
+    else:
+        print("Note: Text was not truncated at sentence boundary")
+
+
+if __name__ == "__main__":
+    demonstrate_truncation()
\ No newline at end of file

From 9576957adf7934e1d3bc5ddf0f8c0deef962f973 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E7=8E=8B=E7=A5=A5=E5=AE=87?= <625024108@qq.com>
Date: Sat, 2 Aug 2025 02:17:48 +0800
Subject: [PATCH 5/5] update

---
 examples/utils/misc/conversation_test_truncate.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/examples/utils/misc/conversation_test_truncate.py b/examples/utils/misc/conversation_test_truncate.py
index ae34692b..f660ff45 100644
--- a/examples/utils/misc/conversation_test_truncate.py
+++ b/examples/utils/misc/conversation_test_truncate.py
@@ -1,6 +1,6 @@
 from swarms.structs.conversation import Conversation
 from dotenv import load_dotenv
-
+from swarms.utils.litellm_tokenizer import count_tokens
 
 # Load environment variables from .env file
 load_dotenv()
@@ -22,7 +22,7 @@ def demonstrate_truncation():
     conversation.add("user", short_message)
     
     # Display token count
-    from swarms.utils.litellm_tokenizer import count_tokens
+    
     tokens = count_tokens(short_message, conversation.tokenizer_model_name)
     print(f"Short message token count: {tokens}")