wx44wx/pipecat
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update "natural conversation" examples to use universal LLMContext

Browse Source
This commit is contained in:
Paul Kompfner
2025-09-23 15:37:58 -04:00
parent 677f69971c
commit 6ccbfd9b57
2 changed files with 58 additions and 72 deletions
Download Patch File Download Diff File Expand all files Collapse all files

65
examples/foundational/22b-natural-conversation-proposal.py
View File

@@ -9,8 +9,9 @@ import os
from dotenv import load_dotenv from dotenv import load_dotenv
from loguru import logger from loguru import logger
from openai.types.chat import ChatCompletionToolParam
from pipecat.adapters.schemas.function_schema import FunctionSchema
from pipecat.adapters.schemas.tools_schema import ToolsSchema
from pipecat.audio.vad.silero import SileroVADAnalyzer from pipecat.audio.vad.silero import SileroVADAnalyzer
from pipecat.frames.frames import ( from pipecat.frames.frames import (
CancelFrame, CancelFrame,
@@ -19,6 +20,7 @@ from pipecat.frames.frames import (
FunctionCallInProgressFrame, FunctionCallInProgressFrame,
FunctionCallResultFrame, FunctionCallResultFrame,
InterruptionFrame, InterruptionFrame,
LLMContextFrame,
LLMRunFrame, LLMRunFrame,
StartFrame, StartFrame,
SystemFrame, SystemFrame,
@@ -32,10 +34,8 @@ from pipecat.pipeline.parallel_pipeline import ParallelPipeline
from pipecat.pipeline.pipeline import Pipeline from pipecat.pipeline.pipeline import Pipeline
from pipecat.pipeline.runner import PipelineRunner from pipecat.pipeline.runner import PipelineRunner
from pipecat.pipeline.task import PipelineParams, PipelineTask from pipecat.pipeline.task import PipelineParams, PipelineTask
from pipecat.processors.aggregators.openai_llm_context import ( from pipecat.processors.aggregators.llm_context import LLMContext
OpenAILLMContext, from pipecat.processors.aggregators.llm_response_universal import LLMContextAggregatorPair
OpenAILLMContextFrame,
)
from pipecat.processors.filters.function_filter import FunctionFilter from pipecat.processors.filters.function_filter import FunctionFilter
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
from pipecat.processors.user_idle_processor import UserIdleProcessor from pipecat.processors.user_idle_processor import UserIdleProcessor
@@ -66,13 +66,13 @@ class StatementJudgeContextFilter(FrameProcessor):
await self.push_frame(frame, direction) await self.push_frame(frame, direction)
return return
# We only want to handle OpenAILLMContextFrames, and only want to push through a simplified # We only want to handle LLMContextFrames, and only want to push through a simplified
# context frame that contains a system prompt and the most recent user messages, # context frame that contains a system prompt and the most recent user messages,
# concatenated. # concatenated.
if isinstance(frame, OpenAILLMContextFrame): if isinstance(frame, LLMContextFrame):
logger.debug(f"Context Frame: {frame}") logger.debug(f"Context Frame: {frame}")
# Take text content from the most recent user messages. # Take text content from the most recent user messages.
messages = frame.context.messages messages = frame.context.get_messages()
user_text_messages = [] user_text_messages = []
last_assistant_message = None last_assistant_message = None
for message in reversed(messages): for message in reversed(messages):
@@ -100,7 +100,7 @@ class StatementJudgeContextFilter(FrameProcessor):
if last_assistant_message: if last_assistant_message:
messages.append(last_assistant_message) messages.append(last_assistant_message)
messages.append({"role": "user", "content": user_message}) messages.append({"role": "user", "content": user_message})
await self.push_frame(OpenAILLMContextFrame(OpenAILLMContext(messages))) await self.push_frame(LLMContextFrame(LLMContext(messages)))
class CompletenessCheck(FrameProcessor): class CompletenessCheck(FrameProcessor):
@@ -231,7 +231,7 @@ class TurnDetectionLLM(Pipeline):
async def pass_only_llm_trigger_frames(frame): async def pass_only_llm_trigger_frames(frame):
return ( return (
isinstance(frame, OpenAILLMContextFrame) isinstance(frame, LLMContextFrame)
or isinstance(frame, InterruptionFrame) or isinstance(frame, InterruptionFrame)
or isinstance(frame, FunctionCallInProgressFrame) or isinstance(frame, FunctionCallInProgressFrame)
or isinstance(frame, FunctionCallResultFrame) or isinstance(frame, FunctionCallResultFrame)
@@ -244,7 +244,7 @@ class TurnDetectionLLM(Pipeline):
[ [
ParallelPipeline( ParallelPipeline(
[ [
# Ignore everything except an OpenAILLMContextFrame. Pass a specially constructed # Ignore everything except an LLMContextFrame. Pass a specially constructed
# simplified context frame to the statement classifier LLM. The only frame this # simplified context frame to the statement classifier LLM. The only frame this
# sub-pipeline will output is a UserStoppedSpeakingFrame. # sub-pipeline will output is a UserStoppedSpeakingFrame.
statement_judge_context_filter, statement_judge_context_filter,
@@ -306,30 +306,23 @@ async def run_bot(transport: BaseTransport, runner_args: RunnerArguments):
async def on_function_calls_started(service, function_calls): async def on_function_calls_started(service, function_calls):
await tts.queue_frame(TTSSpeakFrame("Let me check on that.")) await tts.queue_frame(TTSSpeakFrame("Let me check on that."))
tools = [ weather_function = FunctionSchema(
ChatCompletionToolParam( name="get_current_weather",
type="function", description="Get the current weather",
function={ properties={
"name": "get_current_weather", "location": {
"description": "Get the current weather", "type": "string",
"parameters": { "description": "The city and state, e.g. San Francisco, CA",
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The city and state, e.g. San Francisco, CA",
},
"format": {
"type": "string",
"enum": ["celsius", "fahrenheit"],
"description": "The temperature unit to use. Infer this from the users location.",
},
},
"required": ["location", "format"],
},
}, },
) "format": {
] "type": "string",
"enum": ["celsius", "fahrenheit"],
"description": "The temperature unit to use. Infer this from the users location.",
},
},
required=["location", "format"],
)
tools = ToolsSchema(standard_tools=[weather_function])
messages = [ messages = [
{ {
@@ -338,8 +331,8 @@ async def run_bot(transport: BaseTransport, runner_args: RunnerArguments):
}, },
] ]
context = OpenAILLMContext(messages, tools) context = LLMContext(messages, tools)
context_aggregator = llm_main.create_context_aggregator(context) context_aggregator = LLMContextAggregatorPair(context)
# LLM + turn detection (with an extra LLM as a judge) # LLM + turn detection (with an extra LLM as a judge)
llm = TurnDetectionLLM(llm_main) llm = TurnDetectionLLM(llm_main)

65
examples/foundational/22c-natural-conversation-mixed-llms.py
View File

@@ -9,8 +9,9 @@ import os
from dotenv import load_dotenv from dotenv import load_dotenv
from loguru import logger from loguru import logger
from openai.types.chat import ChatCompletionToolParam
from pipecat.adapters.schemas.function_schema import FunctionSchema
from pipecat.adapters.schemas.tools_schema import ToolsSchema
from pipecat.audio.vad.silero import SileroVADAnalyzer from pipecat.audio.vad.silero import SileroVADAnalyzer
from pipecat.frames.frames import ( from pipecat.frames.frames import (
CancelFrame, CancelFrame,
@@ -19,6 +20,7 @@ from pipecat.frames.frames import (
FunctionCallInProgressFrame, FunctionCallInProgressFrame,
FunctionCallResultFrame, FunctionCallResultFrame,
InterruptionFrame, InterruptionFrame,
LLMContextFrame,
LLMRunFrame, LLMRunFrame,
StartFrame, StartFrame,
SystemFrame, SystemFrame,
@@ -32,10 +34,8 @@ from pipecat.pipeline.parallel_pipeline import ParallelPipeline
from pipecat.pipeline.pipeline import Pipeline from pipecat.pipeline.pipeline import Pipeline
from pipecat.pipeline.runner import PipelineRunner from pipecat.pipeline.runner import PipelineRunner
from pipecat.pipeline.task import PipelineParams, PipelineTask from pipecat.pipeline.task import PipelineParams, PipelineTask
from pipecat.processors.aggregators.openai_llm_context import ( from pipecat.processors.aggregators.llm_context import LLMContext
OpenAILLMContext, from pipecat.processors.aggregators.llm_response_universal import LLMContextAggregatorPair
OpenAILLMContextFrame,
)
from pipecat.processors.filters.function_filter import FunctionFilter from pipecat.processors.filters.function_filter import FunctionFilter
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
from pipecat.processors.user_idle_processor import UserIdleProcessor from pipecat.processors.user_idle_processor import UserIdleProcessor
@@ -272,11 +272,11 @@ class StatementJudgeContextFilter(FrameProcessor):
await self.push_frame(frame, direction) await self.push_frame(frame, direction)
return return
# We only want to handle OpenAILLMContextFrames, and only want to push through a simplified # We only want to handle LLMContextFrames, and only want to push through a simplified
# context frame that contains a system prompt and the most recent user messages, # context frame that contains a system prompt and the most recent user messages,
if isinstance(frame, OpenAILLMContextFrame): if isinstance(frame, LLMContextFrame):
# Take text content from the most recent user messages. # Take text content from the most recent user messages.
messages = frame.context.messages messages = frame.context.get_messages()
user_text_messages = [] user_text_messages = []
last_assistant_message = None last_assistant_message = None
for message in reversed(messages): for message in reversed(messages):
@@ -303,7 +303,7 @@ class StatementJudgeContextFilter(FrameProcessor):
if last_assistant_message: if last_assistant_message:
messages.append(last_assistant_message) messages.append(last_assistant_message)
messages.append({"role": "user", "content": user_message}) messages.append({"role": "user", "content": user_message})
await self.push_frame(OpenAILLMContextFrame(OpenAILLMContext(messages))) await self.push_frame(LLMContextFrame(LLMContext(messages)))
class CompletenessCheck(FrameProcessor): class CompletenessCheck(FrameProcessor):
@@ -425,7 +425,7 @@ class TurnDetectionLLM(Pipeline):
async def pass_only_llm_trigger_frames(frame): async def pass_only_llm_trigger_frames(frame):
return ( return (
isinstance(frame, OpenAILLMContextFrame) isinstance(frame, LLMContextFrame)
or isinstance(frame, InterruptionFrame) or isinstance(frame, InterruptionFrame)
or isinstance(frame, FunctionCallInProgressFrame) or isinstance(frame, FunctionCallInProgressFrame)
or isinstance(frame, FunctionCallResultFrame) or isinstance(frame, FunctionCallResultFrame)
@@ -443,7 +443,7 @@ class TurnDetectionLLM(Pipeline):
FunctionFilter(filter=block_user_stopped_speaking), FunctionFilter(filter=block_user_stopped_speaking),
], ],
[ [
# Ignore everything except an OpenAILLMContextFrame. Pass a specially constructed # Ignore everything except an LLMContextFrame. Pass a specially constructed
# simplified context frame to the statement classifier LLM. The only frame this # simplified context frame to the statement classifier LLM. The only frame this
# sub-pipeline will output is a UserStoppedSpeakingFrame. # sub-pipeline will output is a UserStoppedSpeakingFrame.
statement_judge_context_filter, statement_judge_context_filter,
@@ -509,30 +509,23 @@ async def run_bot(transport: BaseTransport, runner_args: RunnerArguments):
async def on_function_calls_started(service, function_calls): async def on_function_calls_started(service, function_calls):
await tts.queue_frame(TTSSpeakFrame("Let me check on that.")) await tts.queue_frame(TTSSpeakFrame("Let me check on that."))
tools = [ weather_function = FunctionSchema(
ChatCompletionToolParam( name="get_current_weather",
type="function", description="Get the current weather",
function={ properties={
"name": "get_current_weather", "location": {
"description": "Get the current weather", "type": "string",
"parameters": { "description": "The city and state, e.g. San Francisco, CA",
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The city and state, e.g. San Francisco, CA",
},
"format": {
"type": "string",
"enum": ["celsius", "fahrenheit"],
"description": "The temperature unit to use. Infer this from the users location.",
},
},
"required": ["location", "format"],
},
}, },
) "format": {
] "type": "string",
"enum": ["celsius", "fahrenheit"],
"description": "The temperature unit to use. Infer this from the users location.",
},
},
required=["location", "format"],
)
tools = ToolsSchema(standard_tools=[weather_function])
messages = [ messages = [
{ {
@@ -541,8 +534,8 @@ async def run_bot(transport: BaseTransport, runner_args: RunnerArguments):
}, },
] ]
context = OpenAILLMContext(messages, tools) context = LLMContext(messages, tools)
context_aggregator = llm_main.create_context_aggregator(context) context_aggregator = LLMContextAggregatorPair(context)
# LLM + turn detection (with an extra LLM as a judge) # LLM + turn detection (with an extra LLM as a judge)
llm = TurnDetectionLLM(llm_main) llm = TurnDetectionLLM(llm_main)