Don't split bot messages in transcript (Fixes #319)

tests/streaming/agent/test_utils.py

@@ -1,9 +1,25 @@
 from typing import Any, Dict, List, Optional, Union
 from openai.openai_object import OpenAIObject
 from pydantic import BaseModel
-from vocode.streaming.models.actions import FunctionCall
+from vocode.streaming.models.actions import (
+    ActionConfig,
+    ActionInput,
+    ActionOutput,
+    FunctionCall,
+)
 import pytest
-from vocode.streaming.agent.utils import collate_response_async, openai_get_tokens
+from vocode.streaming.agent.utils import (
+    collate_response_async,
+    format_openai_chat_messages_from_transcript,
+    openai_get_tokens,
+)
+from vocode.streaming.models.events import Sender
+from vocode.streaming.models.transcript import (
+    ActionFinish,
+    ActionStart,
+    Message,
+    Transcript,
+)
 
 
 async def _agen_from_list(l):
@@ -202,7 +218,7 @@ class StreamOpenAIResponseTestCase(BaseModel):
             "finish_reason": None,
         },
         {"delta": {}, "finish_reason": "function_call"},
-    ]
+    ],
 ]
 
 EXPECTED_SENTENCES = [
@@ -230,7 +246,7 @@ class StreamOpenAIResponseTestCase(BaseModel):
     ],
     [
         FunctionCall(name="wave_hello", arguments='{\n  "name": "user"\n}'),
-    ]
+    ],
 ]
 
 FUNCTIONS_INPUT = [
@@ -267,7 +283,7 @@ class StreamOpenAIResponseTestCase(BaseModel):
             "finish_reason": None,
         },
         {"delta": {}, "finish_reason": "function_call"},
-    ]
+    ],
 ]
 
 FUNCTIONS_OUTPUT = [
@@ -278,10 +294,10 @@ class StreamOpenAIResponseTestCase(BaseModel):
     ],
     [
         FunctionCall(name="wave_hello", arguments='{\n  "name": "user"\n}'),
-    ]
-
+    ],
 ]
 
+
 @pytest.mark.asyncio
 async def test_stream_openai_response_async():
     test_cases = [
@@ -290,18 +306,123 @@ async def test_stream_openai_response_async():
                 create_chatgpt_openai_object(**obj) for obj in openai_objects
             ],
             expected_sentences=expected_sentences,
-            get_functions=any(isinstance(item, FunctionCall) for item in expected_sentences)
+            get_functions=any(
+                isinstance(item, FunctionCall) for item in expected_sentences
+            ),
         )
         for openai_objects, expected_sentences in zip(
             OPENAI_OBJECTS, EXPECTED_SENTENCES
         )
     ]
-    
+
     for test_case in test_cases:
         actual_sentences = []
         async for sentence in collate_response_async(
             openai_get_tokens(_agen_from_list(test_case.openai_objects)),
-            get_functions=test_case.get_functions
+            get_functions=test_case.get_functions,
         ):
             actual_sentences.append(sentence)
         assert actual_sentences == test_case.expected_sentences
+
+
+def test_format_openai_chat_messages_from_transcript():
+    test_cases = [
+        (
+            (
+                Transcript(
+                    event_logs=[
+                        Message(sender=Sender.BOT, text="Hello!"),
+                        Message(sender=Sender.BOT, text="How are you doing today?"),
+                        Message(sender=Sender.HUMAN, text="I'm doing well, thanks!"),
+                    ]
+                ),
+                "prompt preamble",
+            ),
+            [
+                {"role": "system", "content": "prompt preamble"},
+                {"role": "assistant", "content": "Hello! How are you doing today?"},
+                {"role": "user", "content": "I'm doing well, thanks!"},
+            ],
+        ),
+        (
+            (
+                Transcript(
+                    event_logs=[
+                        Message(sender=Sender.BOT, text="Hello!"),
+                        Message(sender=Sender.BOT, text="How are you doing today?"),
+                        Message(sender=Sender.HUMAN, text="I'm doing well, thanks!"),
+                    ]
+                ),
+                None,
+            ),
+            [
+                {"role": "assistant", "content": "Hello! How are you doing today?"},
+                {"role": "user", "content": "I'm doing well, thanks!"},
+            ],
+        ),
+        (
+            (
+                Transcript(
+                    event_logs=[
+                        Message(sender=Sender.BOT, text="Hello!"),
+                        Message(sender=Sender.BOT, text="How are you doing today?"),
+                    ]
+                ),
+                "prompt preamble",
+            ),
+            [
+                {"role": "system", "content": "prompt preamble"},
+                {"role": "assistant", "content": "Hello! How are you doing today?"},
+            ],
+        ),
+        (
+            (
+                Transcript(
+                    event_logs=[
+                        Message(sender=Sender.BOT, text="Hello!"),
+                        Message(
+                            sender=Sender.HUMAN, text="Hello, what's the weather like?"
+                        ),
+                        ActionStart(
+                            action_type="weather",
+                            action_input=ActionInput(
+                                action_config=ActionConfig(),
+                                conversation_id="asdf",
+                                params={},
+                            ),
+                        ),
+                        ActionFinish(
+                            action_type="weather",
+                            action_output=ActionOutput(
+                                action_type="weather", response={}
+                            ),
+                        ),
+                    ]
+                ),
+                None,
+            ),
+            [
+                {"role": "assistant", "content": "Hello!"},
+                {
+                    "role": "user",
+                    "content": "Hello, what's the weather like?",
+                },
+                {
+                    "role": "assistant",
+                    "content": None,
+                    "function_call": {
+                        "name": "weather",
+                        "arguments": "{}",
+                    },
+                },
+                {
+                    "role": "function",
+                    "name": "weather",
+                    "content": "{}",
+                },
+            ],
+        ),
+    ]
+
+    for params, expected_output in test_cases:
+        assert format_openai_chat_messages_from_transcript(*params) == expected_output

vocode/streaming/agent/utils.py

@@ -1,3 +1,4 @@
+from copy import deepcopy
 import re
 from typing import (
     Dict,
@@ -18,6 +19,7 @@
 from vocode.streaming.models.transcript import (
     ActionFinish,
     ActionStart,
+    EventLog,
     Message,
     Transcript,
 )
@@ -116,7 +118,31 @@ def format_openai_chat_messages_from_transcript(
     chat_messages: List[Dict[str, Optional[Any]]] = (
         [{"role": "system", "content": prompt_preamble}] if prompt_preamble else []
     )
-    for event_log in transcript.event_logs:
+
+    # merge consecutive bot messages
+    new_event_logs: List[EventLog] = []
+    idx = 0
+    while idx < len(transcript.event_logs):
+        bot_messages_buffer: List[Message] = []
+        current_log = transcript.event_logs[idx]
+        while isinstance(current_log, Message) and current_log.sender == Sender.BOT:
+            bot_messages_buffer.append(current_log)
+            idx += 1
+            try:
+                current_log = transcript.event_logs[idx]
+            except IndexError:
+                break
+        if bot_messages_buffer:
+            merged_bot_message = deepcopy(bot_messages_buffer[-1])
+            merged_bot_message.text = " ".join(
+                event_log.text for event_log in bot_messages_buffer
+            )
+            new_event_logs.append(merged_bot_message)
+        else:
+            new_event_logs.append(current_log)
+            idx += 1
+
+    for event_log in new_event_logs:
         if isinstance(event_log, Message):
             chat_messages.append(
                 {