Merge pull request #306 from pipecat-ai/aleix/remove-llm-response-start-end-frame

remove LLMResponseStartFrame and LLMResponseEndFrame

CHANGELOG.md

@@ -5,6 +5,28 @@ All notable changes to **pipecat** will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [Unreleased]
+
+### Removed
+
+- We remove the `LLMResponseStartFrame` and `LLMResponseEndFrame` frames. These
+  were added in the past to properly handle interruptions for the
+  `LLMAssistantContextAggregator`. But the `LLMContextAggregator` is now based
+  on `LLMResponseAggregator` which handles interruptions properly by just
+  processing the `StartInterruptionFrame`, so there's no need for these extra
+  frames any more.
+
+### Fixed
+
+- `TTSService` end of sentence detection has been improved. It now works with
+  acronyms, numbers, hours and others.
+
+### Performance
+
+- `CartesiaTTSService` now uses websockets which improves speed. It also
+  leverages the new Cartesia contexts which maintains generated audio prosody
+  when multiple inputs are sent, therefore improving audio quality a lot.
+
 ## [0.0.36] - 2024-07-02
 
 ### Added

src/pipecat/frames/frames.py

@@ -282,27 +282,13 @@ class EndFrame(ControlFrame):
 
 @dataclass
 class LLMFullResponseStartFrame(ControlFrame):
-    """Used to indicate the beginning of a full LLM response. Following
+    """Used to indicate the beginning of an LLM response. Following by one or
-    LLMResponseStartFrame, TextFrame and LLMResponseEndFrame for each sentence
+    more TextFrame and a final LLMFullResponseEndFrame."""
-    until a LLMFullResponseEndFrame."""
     pass
 
 
 @dataclass
 class LLMFullResponseEndFrame(ControlFrame):
-    """Indicates the end of a full LLM response."""
-    pass
-
-
-@dataclass
-class LLMResponseStartFrame(ControlFrame):
-    """Used to indicate the beginning of an LLM response. Following TextFrames
-    are part of the LLM response until an LLMResponseEndFrame"""
-    pass
-
-
-@dataclass
-class LLMResponseEndFrame(ControlFrame):
     """Indicates the end of an LLM response."""
     pass
 

src/pipecat/processors/aggregators/llm_response.py

@@ -14,8 +14,6 @@ from pipecat.frames.frames import (
     InterimTranscriptionFrame,
     LLMFullResponseEndFrame,
     LLMFullResponseStartFrame,
-    LLMResponseEndFrame,
-    LLMResponseStartFrame,
     LLMMessagesFrame,
     StartInterruptionFrame,
     TranscriptionFrame,
@@ -173,7 +171,7 @@ class LLMUserResponseAggregator(LLMResponseAggregator):
 
 class LLMFullResponseAggregator(FrameProcessor):
     """This class aggregates Text frames until it receives a
-    LLMResponseEndFrame, then emits the concatenated text as
+    LLMFullResponseEndFrame, then emits the concatenated text as
     a single text frame.
 
     given the following frames:
@@ -182,12 +180,12 @@ class LLMFullResponseAggregator(FrameProcessor):
         TextFrame(" world.")
         TextFrame(" I am")
         TextFrame(" an LLM.")
-        LLMResponseEndFrame()]
+        LLMFullResponseEndFrame()]
 
     this processor will yield nothing for the first 4 frames, then
 
         TextFrame("Hello, world. I am an LLM.")
-        LLMResponseEndFrame()
+        LLMFullResponseEndFrame()
 
     when passed the last frame.
 
@@ -203,9 +201,9 @@ class LLMFullResponseAggregator(FrameProcessor):
     >>> asyncio.run(print_frames(aggregator, TextFrame(" world.")))
     >>> asyncio.run(print_frames(aggregator, TextFrame(" I am")))
     >>> asyncio.run(print_frames(aggregator, TextFrame(" an LLM.")))
-    >>> asyncio.run(print_frames(aggregator, LLMResponseEndFrame()))
+    >>> asyncio.run(print_frames(aggregator, LLMFullResponseEndFrame()))
     Hello, world. I am an LLM.
-    LLMResponseEndFrame
+    LLMFullResponseEndFrame
     """
 
     def __init__(self):
@@ -234,6 +232,11 @@ class LLMContextAggregator(LLMResponseAggregator):
     async def _push_aggregation(self):
         if len(self._aggregation) > 0:
             self._context.add_message({"role": self._role, "content": self._aggregation})
+
+            # Reset the aggregation. Reset it before pushing it down, otherwise
+            # if the tasks gets cancelled we won't be able to clear things up.
+            self._aggregation = ""
+
             frame = OpenAILLMContextFrame(self._context)
             await self.push_frame(frame)
 
@@ -247,9 +250,10 @@ class LLMAssistantContextAggregator(LLMContextAggregator):
             messages=[],
             context=context,
             role="assistant",
-            start_frame=LLMResponseStartFrame,
+            start_frame=LLMFullResponseStartFrame,
-            end_frame=LLMResponseEndFrame,
+            end_frame=LLMFullResponseEndFrame,
-            accumulator_frame=TextFrame
+            accumulator_frame=TextFrame,
+            handle_interruptions=True
         )
 
 

src/pipecat/processors/frameworks/langchain.py

@@ -11,8 +11,6 @@ from pipecat.frames.frames import (
     LLMFullResponseEndFrame,
     LLMFullResponseStartFrame,
     LLMMessagesFrame,
-    LLMResponseEndFrame,
-    LLMResponseStartFrame,
     TextFrame)
 from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
 
@@ -69,9 +67,7 @@ class LangchainProcessor(FrameProcessor):
                 {self._transcript_key: text},
                 config={"configurable": {"session_id": self._participant_id}},
             ):
-                await self.push_frame(LLMResponseStartFrame())
                 await self.push_frame(TextFrame(self.__get_token_value(token)))
-                await self.push_frame(LLMResponseEndFrame())
         except GeneratorExit:
             logger.warning(f"{self} generator was closed prematurely")
         except Exception as e:

src/pipecat/services/anthropic.py

@@ -12,8 +12,6 @@ from pipecat.frames.frames import (
     VisionImageRawFrame,
     LLMMessagesFrame,
     LLMFullResponseStartFrame,
-    LLMResponseStartFrame,
-    LLMResponseEndFrame,
     LLMFullResponseEndFrame
 )
 from pipecat.processors.frame_processor import FrameDirection
@@ -118,9 +116,7 @@ class AnthropicLLMService(LLMService):
             async for event in response:
                 # logger.debug(f"Anthropic LLM event: {event}")
                 if (event.type == "content_block_delta"):
-                    await self.push_frame(LLMResponseStartFrame())
                     await self.push_frame(TextFrame(event.delta.text))
-                    await self.push_frame(LLMResponseEndFrame())
 
         except Exception as e:
             logger.exception(f"{self} exception: {e}")

src/pipecat/services/google.py

@@ -14,8 +14,6 @@ from pipecat.frames.frames import (
     VisionImageRawFrame,
     LLMMessagesFrame,
     LLMFullResponseStartFrame,
-    LLMResponseStartFrame,
-    LLMResponseEndFrame,
     LLMFullResponseEndFrame
 )
 from pipecat.processors.frame_processor import FrameDirection
@@ -95,9 +93,7 @@ class GoogleLLMService(LLMService):
             async for chunk in self._async_generator_wrapper(response):
                 try:
                     text = chunk.text
-                    await self.push_frame(LLMResponseStartFrame())
                     await self.push_frame(TextFrame(text))
-                    await self.push_frame(LLMResponseEndFrame())
                 except Exception as e:
                     # Google LLMs seem to flag safety issues a lot!
                     if chunk.candidates[0].finish_reason == 3:

src/pipecat/services/openai.py

@@ -21,8 +21,6 @@ from pipecat.frames.frames import (
     LLMFullResponseEndFrame,
     LLMFullResponseStartFrame,
     LLMMessagesFrame,
-    LLMResponseEndFrame,
-    LLMResponseStartFrame,
     TextFrame,
     URLImageRawFrame,
     VisionImageRawFrame
@@ -151,9 +149,7 @@ class BaseOpenAILLMService(LLMService):
                     # Keep iterating through the response to collect all the argument fragments
                     arguments += tool_call.function.arguments
             elif chunk.choices[0].delta.content:
-                await self.push_frame(LLMResponseStartFrame())
                 await self.push_frame(TextFrame(chunk.choices[0].delta.content))
-                await self.push_frame(LLMResponseEndFrame())
 
         # if we got a function name and arguments, check to see if it's a function with
         # a registered handler. If so, run the registered callback, save the result to

tests/integration/integration_openai_llm.py

@@ -8,8 +8,6 @@ from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
 from pipecat.frames.frames import (
     LLMFullResponseStartFrame,
     LLMFullResponseEndFrame,
-    LLMResponseEndFrame,
-    LLMResponseStartFrame,
     TextFrame
 )
 from pipecat.utils.test_frame_processor import TestFrameProcessor
@@ -64,7 +62,7 @@ if __name__ == "__main__":
         llm.register_function("get_current_weather", get_weather_from_api)
         t = TestFrameProcessor([
             LLMFullResponseStartFrame,
-            [LLMResponseStartFrame, TextFrame, LLMResponseEndFrame],
+            TextFrame,
             LLMFullResponseEndFrame
         ])
         llm.link(t)
@@ -98,7 +96,7 @@ if __name__ == "__main__":
         llm.register_function("get_current_weather", get_weather_from_api)
         t = TestFrameProcessor([
             LLMFullResponseStartFrame,
-            [LLMResponseStartFrame, TextFrame, LLMResponseEndFrame],
+            TextFrame,
             LLMFullResponseEndFrame
         ])
         llm.link(t)
@@ -121,7 +119,7 @@ if __name__ == "__main__":
         api_key = os.getenv("OPENAI_API_KEY")
         t = TestFrameProcessor([
             LLMFullResponseStartFrame,
-            [LLMResponseStartFrame, TextFrame, LLMResponseEndFrame],
+            TextFrame,
             LLMFullResponseEndFrame
         ])
         llm = OpenAILLMService(