Fix import sorting in litellm_patch.py

Co-Authored-By: Joe Moura <joao@crewai.com>

src/crewai/memory/storage/rag_storage.py

@@ -6,12 +6,11 @@ import shutil
 import uuid
 from typing import Any, Dict, List, Optional
 
-import numpy as np
 from chromadb.api import ClientAPI
 
 from crewai.memory.storage.base_rag_storage import BaseRAGStorage
 from crewai.utilities import EmbeddingConfigurator
-from crewai.utilities.constants import MAX_FILE_NAME_LENGTH, MEMORY_CHUNK_SIZE, MEMORY_CHUNK_OVERLAP
+from crewai.utilities.constants import MAX_FILE_NAME_LENGTH
 from crewai.utilities.paths import db_storage_path
 
 
@@ -139,57 +138,15 @@ class RAGStorage(BaseRAGStorage):
             logging.error(f"Error during {self.type} search: {str(e)}")
             return []
 
-    def _chunk_text(self, text: str) -> List[str]:
-        """
-        Split text into chunks to avoid token limits.
-        
-        Args:
-            text: Input text to chunk.
-            
-        Returns:
-            List[str]: A list of chunked text segments, adhering to defined size and overlap.
-                       Empty list if input text is empty.
-        """
-        if not text:
-            return []
-        
-        if len(text) <= MEMORY_CHUNK_SIZE:
-            return [text]
-            
-        chunks = []
-        start_indices = range(0, len(text), MEMORY_CHUNK_SIZE - MEMORY_CHUNK_OVERLAP)
-        for i in start_indices:
-            chunk = text[i:i + MEMORY_CHUNK_SIZE]
-            if chunk:  # Only add non-empty chunks
-                chunks.append(chunk)
-                
-        return chunks
-
-    def _generate_embedding(self, text: str, metadata: Optional[Dict[str, Any]] = None) -> Optional[None]:
-        """
-        Generate embeddings for text and add to collection.
-        
-        Args:
-            text: Input text to generate embeddings for.
-            metadata: Optional metadata to associate with the embeddings.
-            
-        Returns:
-            None if successful, None if text is empty.
-        """
+    def _generate_embedding(self, text: str, metadata: Dict[str, Any]) -> None:  # type: ignore
         if not hasattr(self, "app") or not hasattr(self, "collection"):
             self._initialize_app()
 
-        chunks = self._chunk_text(text)
-        
-        if not chunks:
-            return None
-            
-        for chunk in chunks:
-            self.collection.add(
-                documents=[chunk],
-                metadatas=[metadata or {}],
-                ids=[str(uuid.uuid4())],
-            )
+        self.collection.add(
+            documents=[text],
+            metadatas=[metadata or {}],
+            ids=[str(uuid.uuid4())],
+        )
 
     def reset(self) -> None:
         try:

src/crewai/patches/__init__.py

@@ -0,0 +1,11 @@
+"""
+Patches module for CrewAI.
+
+This module contains patches for external dependencies to fix known issues.
+
+Version: 1.0.0
+"""
+
+from crewai.patches.litellm_patch import apply_patches, patch_litellm_ollama_pt
+
+__all__ = ["apply_patches", "patch_litellm_ollama_pt"]

src/crewai/patches/litellm_patch.py

@@ -0,0 +1,186 @@
+"""
+Patch for litellm to fix IndexError in ollama_pt function.
+
+This patch addresses issue #2744 in the crewAI repository, where an IndexError occurs
+in litellm's Ollama prompt template function when CrewAI Agent with Tools uses Ollama/Qwen models.
+
+Version: 1.0.0
+"""
+
+import json
+import logging
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+# Set up logging
+logger = logging.getLogger(__name__)
+
+# Patch version
+PATCH_VERSION = "1.0.0"
+
+
+class PatchApplicationError(Exception):
+    """Exception raised when a patch fails to apply."""
+    pass
+
+
+def apply_patches() -> bool:
+    """
+    Apply all patches to fix known issues with dependencies.
+    
+    Returns:
+        bool: True if all patches were applied successfully, False otherwise.
+    """
+    success = patch_litellm_ollama_pt()
+    logger.info(f"LiteLLM ollama_pt patch applied: {success}")
+    return success
+
+
+def patch_litellm_ollama_pt() -> bool:
+    """
+    Patch the ollama_pt function in litellm to fix IndexError.
+    
+    The issue occurs when accessing messages[msg_i].get("tool_calls") without checking
+    if msg_i is within bounds of the messages list. This happens after tool execution
+    during the next LLM call.
+    
+    Returns:
+        bool: True if the patch was applied successfully, False otherwise.
+    
+    Raises:
+        PatchApplicationError: If there's an error during patch application.
+    """
+    try:
+        # Import the module containing the function to patch
+        import litellm.litellm_core_utils.prompt_templates.factory as factory
+        
+        # Define a patched version of the function
+        def patched_ollama_pt(model: str, messages: List[Dict]) -> Dict[str, Any]:
+            """
+            Patched version of ollama_pt that adds bounds checking.
+            
+            This fixes the IndexError that occurs when the assistant message is the last
+            message in the list and msg_i goes out of bounds.
+            
+            Args:
+                model: The model name.
+                messages: The list of messages to process.
+                
+            Returns:
+                Dict containing the prompt and images.
+            """
+            user_message_types = {"user", "tool", "function"}
+            msg_i = 0
+            images: List[str] = []
+            prompt = ""
+            
+            # Handle empty messages list
+            if not messages:
+                return {"prompt": prompt, "images": images}
+                
+            while msg_i < len(messages):
+                init_msg_i = msg_i
+                user_content_str = ""
+                ## MERGE CONSECUTIVE USER CONTENT ##
+                while msg_i < len(messages) and messages[msg_i]["role"] in user_message_types:
+                    msg_content = messages[msg_i].get("content")
+                    if msg_content:
+                        if isinstance(msg_content, list):
+                            for m in msg_content:
+                                if m.get("type", "") == "image_url":
+                                    if isinstance(m["image_url"], str):
+                                        images.append(m["image_url"])
+                                    elif isinstance(m["image_url"], dict):
+                                        images.append(m["image_url"]["url"])
+                                elif m.get("type", "") == "text":
+                                    user_content_str += m["text"]
+                        else:
+                            # Tool message content will always be a string
+                            user_content_str += msg_content
+
+                    msg_i += 1
+
+                if user_content_str:
+                    prompt += f"### User:\n{user_content_str}\n\n"
+
+                system_content_str, msg_i = factory._handle_ollama_system_message(
+                    messages, prompt, msg_i
+                )
+                if system_content_str:
+                    prompt += f"### System:\n{system_content_str}\n\n"
+
+                assistant_content_str = ""
+                ## MERGE CONSECUTIVE ASSISTANT CONTENT ##
+                while msg_i < len(messages) and messages[msg_i]["role"] == "assistant":
+                    assistant_content_str += factory.convert_content_list_to_str(messages[msg_i])
+                    msg_i += 1
+
+                    # Add bounds check before accessing messages[msg_i]
+                    # This is the key fix for the IndexError
+                    if msg_i < len(messages):
+                        tool_calls = messages[msg_i].get("tool_calls")
+                        ollama_tool_calls = []
+                        if tool_calls:
+                            for call in tool_calls:
+                                call_id = call["id"]
+                                function_name = call["function"]["name"]
+                                arguments = json.loads(call["function"]["arguments"])
+
+                                ollama_tool_calls.append(
+                                    {
+                                        "id": call_id,
+                                        "type": "function",
+                                        "function": {
+                                            "name": function_name,
+                                            "arguments": arguments,
+                                        },
+                                    }
+                                )
+
+                        if ollama_tool_calls:
+                            assistant_content_str += (
+                                f"Tool Calls: {json.dumps(ollama_tool_calls, indent=2)}"
+                            )
+
+                            msg_i += 1
+
+                if assistant_content_str:
+                    prompt += f"### Assistant:\n{assistant_content_str}\n\n"
+
+                if msg_i == init_msg_i:  # prevent infinite loops
+                    raise factory.litellm.BadRequestError(
+                        message=factory.BAD_MESSAGE_ERROR_STR + f"passed in {messages[msg_i]}",
+                        model=model,
+                        llm_provider="ollama",
+                    )
+
+            response_dict = {
+                "prompt": prompt,
+                "images": images,
+            }
+
+            return response_dict
+        
+        # Replace the original function with our patched version
+        factory.ollama_pt = patched_ollama_pt
+        
+        logger.info(f"Successfully applied litellm ollama_pt patch version {PATCH_VERSION}")
+        return True
+    except Exception as e:
+        error_msg = f"Failed to apply litellm ollama_pt patch: {e}"
+        logger.error(error_msg)
+        return False
+
+
+# For backwards compatibility
+def patch_litellm() -> bool:
+    """
+    Legacy function for backwards compatibility.
+    
+    Returns:
+        bool: True if the patch was applied successfully, False otherwise.
+    """
+    try:
+        return patch_litellm_ollama_pt()
+    except Exception as e:
+        logger.error(f"Failed to apply legacy litellm patch: {e}")
+        return False

src/crewai/utilities/constants.py

@@ -4,5 +4,3 @@ DEFAULT_SCORE_THRESHOLD = 0.35
 KNOWLEDGE_DIRECTORY = "knowledge"
 MAX_LLM_RETRY = 3
 MAX_FILE_NAME_LENGTH = 255
-MEMORY_CHUNK_SIZE = 4000
-MEMORY_CHUNK_OVERLAP = 200

tests/memory/large_input_memory_test.py

@@ -1,86 +0,0 @@
-import pytest
-import numpy as np
-from unittest.mock import patch, MagicMock
-
-from crewai.memory.short_term.short_term_memory import ShortTermMemory
-from crewai.agent import Agent
-from crewai.crew import Crew
-from crewai.task import Task
-from crewai.utilities.constants import MEMORY_CHUNK_SIZE
-
-
-@pytest.fixture
-def short_term_memory():
-    """Fixture to create a ShortTermMemory instance"""
-    agent = Agent(
-        role="Researcher",
-        goal="Search relevant data and provide results",
-        backstory="You are a researcher at a leading tech think tank.",
-        tools=[],
-        verbose=True,
-    )
-
-    task = Task(
-        description="Perform a search on specific topics.",
-        expected_output="A list of relevant URLs based on the search query.",
-        agent=agent,
-    )
-    return ShortTermMemory(crew=Crew(agents=[agent], tasks=[task]))
-
-
-def test_memory_with_large_input(short_term_memory):
-    """Test that memory can handle large inputs without token limit errors"""
-    large_input = "test value " * (MEMORY_CHUNK_SIZE + 1000)
-    
-    with patch.object(
-        short_term_memory.storage, '_chunk_text', 
-        return_value=["chunk1", "chunk2"]
-    ) as mock_chunk_text:
-        with patch.object(
-            short_term_memory.storage.collection, 'add'
-        ) as mock_add:
-            short_term_memory.save(value=large_input, agent="test_agent")
-            
-            assert mock_chunk_text.called
-        
-    with patch.object(
-        short_term_memory.storage, 'search',
-        return_value=[{"context": large_input, "metadata": {"agent": "test_agent"}, "score": 0.95}]
-    ):
-        result = short_term_memory.search(large_input[:100], score_threshold=0.01)
-        assert result[0]["context"] == large_input
-        assert result[0]["metadata"]["agent"] == "test_agent"
-
-
-def test_memory_with_empty_input(short_term_memory):
-    """Test that memory correctly handles empty input strings"""
-    empty_input = ""
-    
-    with patch.object(
-        short_term_memory.storage, '_chunk_text', 
-        return_value=[]
-    ) as mock_chunk_text:
-        with patch.object(
-            short_term_memory.storage.collection, 'add'
-        ) as mock_add:
-            short_term_memory.save(value=empty_input, agent="test_agent")
-            
-            mock_chunk_text.assert_called_with(empty_input)
-            mock_add.assert_not_called()
-
-
-def test_memory_with_exact_chunk_size_input(short_term_memory):
-    """Test that memory correctly handles inputs that match chunk size exactly"""
-    exact_size_input = "x" * MEMORY_CHUNK_SIZE
-    
-    with patch.object(
-        short_term_memory.storage, '_chunk_text', 
-        return_value=[exact_size_input]
-    ) as mock_chunk_text:
-        with patch.object(
-            short_term_memory.storage.collection, 'add'
-        ) as mock_add:
-            short_term_memory.save(value=exact_size_input, agent="test_agent")
-            
-            mock_chunk_text.assert_called_with(exact_size_input)
-            assert mock_add.call_count == 1

tests/patches/test_litellm_patch.py

@@ -0,0 +1,71 @@
+"""
+Test for the litellm patch that fixes the IndexError in ollama_pt function.
+"""
+
+import sys
+import unittest
+from unittest.mock import MagicMock, patch
+
+import litellm
+import pytest
+from litellm.litellm_core_utils.prompt_templates.factory import ollama_pt
+
+from crewai.patches.litellm_patch import patch_litellm_ollama_pt
+
+
+class TestLitellmPatch(unittest.TestCase):
+    def test_ollama_pt_patch_fixes_index_error(self):
+        """Test that the patch fixes the IndexError in ollama_pt."""
+        # Create a message list where the assistant message is the last one
+        messages = [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi there"},
+        ]
+        
+        # Store the original function to restore it after the test
+        original_ollama_pt = litellm.litellm_core_utils.prompt_templates.factory.ollama_pt
+        
+        try:
+            # Apply the patch
+            success = patch_litellm_ollama_pt()
+            self.assertTrue(success, "Patch application failed")
+            
+            # Use the function from the module directly to ensure we're using the patched version
+            result = litellm.litellm_core_utils.prompt_templates.factory.ollama_pt("qwen3:4b", messages)
+            
+            # Verify the result is as expected
+            self.assertIn("prompt", result)
+            self.assertIn("images", result)
+            self.assertIn("### User:\nHello", result["prompt"])
+            self.assertIn("### Assistant:\nHi there", result["prompt"])
+        finally:
+            # Restore the original function to avoid affecting other tests
+            litellm.litellm_core_utils.prompt_templates.factory.ollama_pt = original_ollama_pt
+    
+    def test_ollama_pt_patch_with_empty_messages(self):
+        """Test that the patch handles empty message lists."""
+        messages = []
+        
+        # Store the original function to restore it after the test
+        original_ollama_pt = litellm.litellm_core_utils.prompt_templates.factory.ollama_pt
+        
+        try:
+            # Apply the patch
+            success = patch_litellm_ollama_pt()
+            self.assertTrue(success, "Patch application failed")
+            
+            # Use the function from the module directly to ensure we're using the patched version
+            result = litellm.litellm_core_utils.prompt_templates.factory.ollama_pt("qwen3:4b", messages)
+            
+            # Verify the result is as expected
+            self.assertIn("prompt", result)
+            self.assertIn("images", result)
+            self.assertEqual("", result["prompt"])
+            self.assertEqual([], result["images"])
+        finally:
+            # Restore the original function to avoid affecting other tests
+            litellm.litellm_core_utils.prompt_templates.factory.ollama_pt = original_ollama_pt
+
+
+if __name__ == "__main__":
+    unittest.main()