fix(perplexity): keep tool calls when merging consecutive assistant messages

changelog/5855.fixed.md

@@ -0,0 +1 @@
+- Fixed a bug in the Perplexity adapter where merging consecutive assistant messages dropped `tool_calls` and collapsed consecutive tool messages onto the first call's id.

src/pipecat/adapters/llm/services/perplexity_adapter.py

@@ -151,17 +151,24 @@ def _transform_messages(
             if current["role"] == next_msg["role"] == "system":
                 # Perplexity allows multiple initial system messages, don't merge
                 i += 1
-            elif current["role"] == next_msg["role"]:
+            # A tool message names the call its result answers in tool_call_id,
+            # so merging two of them would leave both results on the first
+            # call's id. Leave consecutive tool messages as they are.
+            elif current["role"] == next_msg["role"] and "tool_call_id" not in next_msg:
                 # Convert string content to list-of-dicts format for merging
                 if isinstance(current.get("content"), str):
                     current["content"] = [{"type": "text", "text": current["content"]}]
                 if isinstance(next_msg.get("content"), str):
                     next_msg["content"] = [{"type": "text", "text": next_msg["content"]}]
                 # Merge content from next message into current
-                if isinstance(current.get("content"), list) and isinstance(
-                    next_msg.get("content"), list
-                ):
-                    current["content"].extend(next_msg["content"])
+                if isinstance(next_msg.get("content"), list):
+                    if isinstance(current.get("content"), list):
+                        current["content"].extend(next_msg["content"])
+                    else:
+                        current["content"] = next_msg["content"]
+                # tool_calls are carried outside "content", so merge those too
+                if next_msg.get("tool_calls"):
+                    current.setdefault("tool_calls", []).extend(next_msg["tool_calls"])
                 msgs.pop(i + 1)
             else:
                 i += 1

tests/test_get_llm_invocation_params.py

@@ -2906,6 +2906,114 @@ async def test_tool_messages_preserved(self):
         self.assertEqual(params["messages"][2]["content"], "Sunny, 72F")
         self.assertEqual(params["messages"][3]["role"], "user")
 
+    async def test_merging_assistants_keeps_tool_calls(self):
+        """Merging a text assistant message with a tool-call-only one keeps the call.
+
+        Some conversation histories carry the assistant's text and its tool call
+        as two consecutive assistant messages; the tool message that follows
+        answers the call carried by the second one.
+        """
+        messages: list[LLMStandardMessage] = [
+            {"role": "user", "content": "Weather in LA?"},
+            {"role": "assistant", "content": "Let me check on that."},
+            {
+                "role": "assistant",
+                "tool_calls": [
+                    {
+                        "id": "call_1",
+                        "type": "function",
+                        "function": {"name": "get_weather", "arguments": '{"location": "LA"}'},
+                    }
+                ],
+            },
+            {"role": "tool", "content": '{"temperature": "75"}', "tool_call_id": "call_1"},
+            {"role": "user", "content": "Thanks!"},
+        ]
+
+        context = LLMContext(messages=messages)
+        params = await self.adapter.get_llm_invocation_params(
+            context, convert_developer_to_user=True
+        )
+
+        self.assertEqual(len(params["messages"]), 4)
+        merged = params["messages"][1]
+        self.assertEqual(merged["role"], "assistant")
+        self.assertEqual([part["text"] for part in merged["content"]], ["Let me check on that."])
+        self.assertEqual([call["id"] for call in merged["tool_calls"]], ["call_1"])
+        self.assertEqual(params["messages"][2]["role"], "tool")
+        self.assertEqual(params["messages"][2]["tool_call_id"], "call_1")
+        self.assertEqual(params["messages"][3]["role"], "user")
+
+    async def test_merging_tool_call_only_assistant_keeps_content(self):
+        """A history can carry the call before the assistant's text.
+
+        Merging then has to move the text into the tool-call message rather
+        than drop it, since the message holding it is the one that is removed.
+        """
+        messages: list[LLMStandardMessage] = [
+            {"role": "user", "content": "Weather in LA?"},
+            {
+                "role": "assistant",
+                "tool_calls": [
+                    {
+                        "id": "call_1",
+                        "type": "function",
+                        "function": {"name": "get_weather", "arguments": '{"location": "LA"}'},
+                    }
+                ],
+            },
+            {"role": "assistant", "content": "Looks sunny."},
+            {"role": "user", "content": "Thanks!"},
+        ]
+
+        context = LLMContext(messages=messages)
+        params = await self.adapter.get_llm_invocation_params(
+            context, convert_developer_to_user=True
+        )
+
+        self.assertEqual(len(params["messages"]), 3)
+        merged = params["messages"][1]
+        self.assertEqual(merged["role"], "assistant")
+        self.assertEqual([call["id"] for call in merged["tool_calls"]], ["call_1"])
+        self.assertEqual([part["text"] for part in merged["content"]], ["Looks sunny."])
+        self.assertEqual(params["messages"][2]["role"], "user")
+
+    async def test_consecutive_tool_messages_keep_their_own_ids(self):
+        """Tool results for different calls are left apart rather than merged."""
+        messages: list[LLMStandardMessage] = [
+            {"role": "user", "content": "Weather in LA and SF?"},
+            {
+                "role": "assistant",
+                "tool_calls": [
+                    {
+                        "id": "call_1",
+                        "type": "function",
+                        "function": {"name": "get_weather", "arguments": '{"location": "LA"}'},
+                    },
+                    {
+                        "id": "call_2",
+                        "type": "function",
+                        "function": {"name": "get_weather", "arguments": '{"location": "SF"}'},
+                    },
+                ],
+            },
+            {"role": "tool", "content": '{"temperature": "75"}', "tool_call_id": "call_1"},
+            {"role": "tool", "content": '{"temperature": "62"}', "tool_call_id": "call_2"},
+            {"role": "user", "content": "Thanks!"},
+        ]
+
+        context = LLMContext(messages=messages)
+        params = await self.adapter.get_llm_invocation_params(
+            context, convert_developer_to_user=True
+        )
+
+        self.assertEqual(len(params["messages"]), 5)
+        self.assertEqual(params["messages"][2]["tool_call_id"], "call_1")
+        self.assertEqual(params["messages"][2]["content"], '{"temperature": "75"}')
+        self.assertEqual(params["messages"][3]["tool_call_id"], "call_2")
+        self.assertEqual(params["messages"][3]["content"], '{"temperature": "62"}')
+        self.assertEqual(params["messages"][4]["role"], "user")
+
     async def test_developer_message_converted_to_user(self):
         """Developer messages are converted to user role."""
         messages: list[LLMStandardMessage] = [