Made the message role of ReAct observation configurable (#17521)

docs/docs/examples/agent/react_agent.ipynb

@@ -53,22 +53,7 @@
    "execution_count": null,
    "id": "e8ac1778-0585-43c9-9dad-014d13d7460d",
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "[nltk_data] Downloading package stopwords to /Users/jerryliu/Programmi\n",
-      "[nltk_data]     ng/gpt_index/.venv/lib/python3.10/site-\n",
-      "[nltk_data]     packages/llama_index/legacy/_static/nltk_cache...\n",
-      "[nltk_data]   Unzipping corpora/stopwords.zip.\n",
-      "[nltk_data] Downloading package punkt to /Users/jerryliu/Programming/g\n",
-      "[nltk_data]     pt_index/.venv/lib/python3.10/site-\n",
-      "[nltk_data]     packages/llama_index/legacy/_static/nltk_cache...\n",
-      "[nltk_data]   Unzipping tokenizers/punkt.zip.\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
     "from llama_index.core.agent import ReActAgent\n",
     "from llama_index.llms.openai import OpenAI\n",
@@ -468,13 +453,44 @@
     "response = agent.chat(\"What is 5+3+2\")\n",
     "print(response)"
    ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "76190511-692c-4642-9b86-adac88c98550",
+   "metadata": {},
+   "source": [
+    "### Customizing the Message Role of Observation\n",
+    "\n",
+    "If the LLM you use supports function/tool calling, you may set the message role of observations to `MessageRole.TOOL`.  \n",
+    "Doing this will prevent the tool outputs from being misinterpreted as new user messages for some models."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e6d5e8c1-c40e-4a96-8d2e-84127f066265",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from llama_index.core.agent import ReActChatFormatter\n",
+    "from llama_index.core.llms import MessageRole\n",
+    "\n",
+    "agent = ReActAgent.from_tools(\n",
+    "    [multiply_tool, add_tool],\n",
+    "    llm=llm,\n",
+    "    react_chat_formatter=ReActChatFormatter.from_defaults(\n",
+    "        observation_role=MessageRole.TOOL\n",
+    "    ),\n",
+    "    verbose=True,\n",
+    ")"
+   ]
   }
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "llama_index_v2",
+   "display_name": "LlamaIndex Development",
    "language": "python",
-   "name": "llama_index_v2"
+   "name": "llama-index-dev"
   },
   "language_info": {
    "codemirror_mode": {

llama-index-core/llama_index/core/agent/react/formatter.py

@@ -13,7 +13,7 @@
     ObservationReasoningStep,
 )
 from llama_index.core.base.llms.types import ChatMessage, MessageRole
-from llama_index.core.bridge.pydantic import BaseModel, ConfigDict
+from llama_index.core.bridge.pydantic import BaseModel, ConfigDict, Field
 from llama_index.core.tools import BaseTool
 
 logger = logging.getLogger(__name__)
@@ -53,6 +53,14 @@ class ReActChatFormatter(BaseAgentChatFormatter):
 
     system_header: str = REACT_CHAT_SYSTEM_HEADER  # default
     context: str = ""  # not needed w/ default
+    observation_role: MessageRole = Field(
+        default=MessageRole.USER,
+        description=(
+            "Message role of tool outputs. If the LLM you use supports function/tool "
+            "calling, you may set it to `MessageRole.TOOL` to avoid the tool outputs "
+            "being misinterpreted as new user messages."
+        ),
+    )
 
     def format(
         self,
@@ -73,13 +81,13 @@ def format(
         fmt_sys_header = self.system_header.format(**format_args)
 
         # format reasoning history as alternating user and assistant messages
-        # where the assistant messages are thoughts and actions and the user
+        # where the assistant messages are thoughts and actions and the tool
         # messages are observations
         reasoning_history = []
         for reasoning_step in current_reasoning:
             if isinstance(reasoning_step, ObservationReasoningStep):
                 message = ChatMessage(
-                    role=MessageRole.USER,
+                    role=self.observation_role,
                     content=reasoning_step.get_content(),
                 )
             else:
@@ -100,6 +108,7 @@ def from_defaults(
         cls,
         system_header: Optional[str] = None,
         context: Optional[str] = None,
+        observation_role: MessageRole = MessageRole.USER,
     ) -> "ReActChatFormatter":
         """Create ReActChatFormatter from defaults."""
         if not system_header:
@@ -112,6 +121,7 @@ def from_defaults(
         return ReActChatFormatter(
             system_header=system_header,
             context=context or "",
+            observation_role=observation_role,
         )
 
     @classmethod