dv-pipecat-ai 0.0.85.dev7__py3-none-any.whl → 0.0.85.dev11__py3-none-any.whl

{dv_pipecat_ai-0.0.85.dev7.dist-info → dv_pipecat_ai-0.0.85.dev11.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dv-pipecat-ai
-Version: 0.0.85.dev7
+Version: 0.0.85.dev11
 Summary: An open source framework for voice (and multimodal) assistants
 License-Expression: BSD-2-Clause
 Project-URL: Source, https://github.com/pipecat-ai/pipecat
@@ -143,6 +143,9 @@ Requires-Dist: coremltools>=8.0; extra == "local-smart-turn"
 Requires-Dist: transformers; extra == "local-smart-turn"
 Requires-Dist: torch<3,>=2.5.0; extra == "local-smart-turn"
 Requires-Dist: torchaudio<3,>=2.5.0; extra == "local-smart-turn"
+Provides-Extra: local-smart-turn-v3
+Requires-Dist: transformers; extra == "local-smart-turn-v3"
+Requires-Dist: onnxruntime<2,>=1.20.1; extra == "local-smart-turn-v3"
 Provides-Extra: remote-smart-turn
 Provides-Extra: silero
 Requires-Dist: onnxruntime~=1.20.1; extra == "silero"

{dv_pipecat_ai-0.0.85.dev7.dist-info → dv_pipecat_ai-0.0.85.dev11.dist-info}/RECORD

@@ -1,4 +1,4 @@
-dv_pipecat_ai-0.0.85.dev7.dist-info/licenses/LICENSE,sha256=DWY2QGf2eMCFhuu2ChairtT6CB7BEFffNVhXWc4Od08,1301
+dv_pipecat_ai-0.0.85.dev11.dist-info/licenses/LICENSE,sha256=DWY2QGf2eMCFhuu2ChairtT6CB7BEFffNVhXWc4Od08,1301
 pipecat/__init__.py,sha256=j0Xm6adxHhd7D06dIyyPV_GlBYLlBnTAERVvD_jAARQ,861
 pipecat/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 pipecat/adapters/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -50,14 +50,17 @@ pipecat/audio/resamplers/resampy_resampler.py,sha256=fEZv6opn_9j50xYEOdwQiZOJQ_J
 pipecat/audio/resamplers/soxr_resampler.py,sha256=CXze7zf_ExlCcgcZp0oArRSbZ9zFpBzsCt2EQ_woKfM,1747
 pipecat/audio/resamplers/soxr_stream_resampler.py,sha256=lHk1__M1HDGf25abpffuWEyqbd0ckNfyADDV_WmTPcY,3665
 pipecat/audio/turn/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-pipecat/audio/turn/base_turn_analyzer.py,sha256=hLOcH1WkP9iSk84boQv94RFYKEfEX-IHfO1y9pjkDzs,3213
+pipecat/audio/turn/base_turn_analyzer.py,sha256=UoZ61yto2wecXU6nXk2yjdcgM7jGyfMR5ZfrunOFpOA,3359
 pipecat/audio/turn/smart_turn/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-pipecat/audio/turn/smart_turn/base_smart_turn.py,sha256=HgUoRfo9tbXVMfmRbYBkm4FDY1AjUJ3CRe7t48Ny2WI,9672
+pipecat/audio/turn/smart_turn/base_smart_turn.py,sha256=gE5jrqrU0gQcgjTOvpUbb6LWAhfk8VKZQ-5pyEIZH4E,10037
 pipecat/audio/turn/smart_turn/fal_smart_turn.py,sha256=neahuTAY9SUQjacRYd19BERiuSHIMSpqzZ9uae_ZlWA,1606
-pipecat/audio/turn/smart_turn/http_smart_turn.py,sha256=s5QP2gd0BqQAlbRJ7hGuCwGqgEENfyRm6aB6jBgDoqE,4642
+pipecat/audio/turn/smart_turn/http_smart_turn.py,sha256=HlHpdVbk-1g_AU3qAAy7Xob8M2V3FUqtr38UAk1F1Dw,4783
 pipecat/audio/turn/smart_turn/local_coreml_smart_turn.py,sha256=50kiBeZhnq7FZWZnzdSX8KUmhhQtkme0KH2rbiAJbCU,3140
-pipecat/audio/turn/smart_turn/local_smart_turn.py,sha256=KVodqUTu8onfmfeOywgH98vBCNvBb-B3pvsQlTKyP_4,3570
-pipecat/audio/turn/smart_turn/local_smart_turn_v2.py,sha256=aYLMDURpmYycQgKsxbNEENtUe5oujeQ9H3Lbi0GYmZA,7160
+pipecat/audio/turn/smart_turn/local_smart_turn.py,sha256=0z2M_MC9xIcelm4d9XqZwzJMe2FM-zOjgnHDAeoMw0g,3564
+pipecat/audio/turn/smart_turn/local_smart_turn_v2.py,sha256=hd_nhEdaxwJ2_G6F2RJru9mC8vyzkmku2YqmtULl7NM,7154
+pipecat/audio/turn/smart_turn/local_smart_turn_v3.py,sha256=x1q437Mp8cEU1S-7W869i1meDtCdjrjPTUCjbSLDVgQ,4649
+pipecat/audio/turn/smart_turn/data/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+pipecat/audio/turn/smart_turn/data/smart-turn-v3.0.onnx,sha256=B6Ezq6MeLQtSPxf4wuTmXv5tj2he_RLKT-Iev055iZE,8757193
 pipecat/audio/vad/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 pipecat/audio/vad/silero.py,sha256=Cz4_hJjaBKbmUwZVbqMzED8orHOCsnF3zpERgBTw1Rw,7906
 pipecat/audio/vad/vad_analyzer.py,sha256=XkZLEe4z7Ja0lGoYZst1HNYqt5qOwG-vjsk_w8chiNA,7430
@@ -286,6 +289,8 @@ pipecat/services/rime/tts.py,sha256=XHMSnQUi7gMtWF42u4rBVv6oBDor4KkwkL7O-Sj9MPo,
 pipecat/services/riva/__init__.py,sha256=rObSsj504O_TMXhPBg_ymqKslZBhovlR-A0aaRZ0O6A,276
 pipecat/services/riva/stt.py,sha256=dtg8toijmexWB3uipw0EQ7ov3DFgHj40kFFv1Zadmmc,25116
 pipecat/services/riva/tts.py,sha256=idbqx3I2NlWCXtrIFsjEaYapxA3BLIA14ai3aMBh-2w,8158
+pipecat/services/salesforce/__init__.py,sha256=OFvYbcvCadYhcKdBAVLj3ZUXVXQ1HyVyhgxIFf6_Thg,173
+pipecat/services/salesforce/llm.py,sha256=mpozkzldgz3plbMOJcKddiyJxn7x4qqPuJVn22a41Ag,23009
 pipecat/services/sambanova/__init__.py,sha256=oTXExLic-qTcsfsiWmssf3Elclf3IIWoN41_2IpoF18,128
 pipecat/services/sambanova/llm.py,sha256=5XVfPLEk__W8ykFqLdV95ZUhlGGkAaJwmbciLdZYtTc,8976
 pipecat/services/sambanova/stt.py,sha256=ZZgEZ7WQjLFHbCko-3LNTtVajjtfUvbtVLtFcaNadVQ,2536
@@ -358,6 +363,7 @@ pipecat/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 pipecat/utils/base_object.py,sha256=62e5_0R_rcQe-JdzUM0h1wtv1okw-0LPyG78ZKkDyzE,5963
 pipecat/utils/logger_config.py,sha256=5-RmvReZIINeqSXz3ALhEIiMZ_azmpOxnlIkdyCjWWk,5606
 pipecat/utils/network.py,sha256=RRQ7MmTcbeXBJ2aY5UbMCQ6elm5B8Rxkn8XqkJ9S0Nc,825
+pipecat/utils/redis.py,sha256=JmBaC1yY6e8qygUQkAER3DNFCYSCH18hd7NN9qqjDMU,1677
 pipecat/utils/string.py,sha256=TskK9KxQSwbljct0J6y9ffkRcx4xYjTtPhFjEL4M1i8,6720
 pipecat/utils/time.py,sha256=lirjh24suz9EI1pf2kYwvAYb3I-13U_rJ_ZRg3nRiGs,1741
 pipecat/utils/utils.py,sha256=T2y1Mcd9yWiZiIToUiRkhW-n7EFf8juk3kWX3TF8XOQ,2451
@@ -378,7 +384,7 @@ pipecat/utils/tracing/service_decorators.py,sha256=HwDCqLGijhYD3F8nxDuQmEw-YkRw0
 pipecat/utils/tracing/setup.py,sha256=7TEgPNpq6M8lww8OQvf0P9FzYc5A30xICGklVA-fua0,2892
 pipecat/utils/tracing/turn_context_provider.py,sha256=ikon3plFOx0XbMrH6DdeHttNpb-U0gzMZIm3bWLc9eI,2485
 pipecat/utils/tracing/turn_trace_observer.py,sha256=dma16SBJpYSOE58YDWy89QzHyQFc_9gQZszKeWixuwc,9725
-dv_pipecat_ai-0.0.85.dev7.dist-info/METADATA,sha256=VKliyHrf__BVRdCbvDCp5lJRyqMFTxC3-EEuhwPphe0,32691
-dv_pipecat_ai-0.0.85.dev7.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-dv_pipecat_ai-0.0.85.dev7.dist-info/top_level.txt,sha256=kQzG20CxGf-nSsHmtXHx3hY2-8zHA3jYg8jk0TajqXc,8
-dv_pipecat_ai-0.0.85.dev7.dist-info/RECORD,,
+dv_pipecat_ai-0.0.85.dev11.dist-info/METADATA,sha256=_scIy5gP8k7GUtLAA9NzNVT_T1y__8ROU0gPj1G6FCw,32858
+dv_pipecat_ai-0.0.85.dev11.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+dv_pipecat_ai-0.0.85.dev11.dist-info/top_level.txt,sha256=kQzG20CxGf-nSsHmtXHx3hY2-8zHA3jYg8jk0TajqXc,8
+dv_pipecat_ai-0.0.85.dev11.dist-info/RECORD,,

pipecat/audio/turn/base_turn_analyzer.py

@@ -14,6 +14,8 @@ from abc import ABC, abstractmethod
 from enum import Enum
 from typing import Optional, Tuple
 
+from pydantic import BaseModel
+
 from pipecat.metrics.metrics import MetricsData
 
 
@@ -29,6 +31,12 @@ class EndOfTurnState(Enum):
     INCOMPLETE = 2
 
 
+class BaseTurnParams(BaseModel):
+    """Base class for turn analyzer parameters."""
+
+    pass
+
+
 class BaseTurnAnalyzer(ABC):
     """Abstract base class for analyzing user end of turn.
 
@@ -78,7 +86,7 @@ class BaseTurnAnalyzer(ABC):
 
     @property
     @abstractmethod
-    def params(self):
+    def params(self) -> BaseTurnParams:
         """Get the current turn analyzer parameters.
 
         Returns:

pipecat/audio/turn/smart_turn/base_smart_turn.py

@@ -11,15 +11,17 @@ machine learning models to determine when a user has finished speaking, going
 beyond simple silence-based detection.
 """
 
+import asyncio
 import time
 from abc import abstractmethod
+from concurrent.futures import ThreadPoolExecutor
 from typing import Any, Dict, Optional, Tuple
 
 import numpy as np
 from loguru import logger
 from pydantic import BaseModel
 
-from pipecat.audio.turn.base_turn_analyzer import BaseTurnAnalyzer, EndOfTurnState
+from pipecat.audio.turn.base_turn_analyzer import BaseTurnAnalyzer, BaseTurnParams, EndOfTurnState
 from pipecat.metrics.metrics import MetricsData, SmartTurnMetricsData
 
 # Default timing parameters
@@ -29,7 +31,7 @@ MAX_DURATION_SECONDS = 8  # Max allowed segment duration
 USE_ONLY_LAST_VAD_SEGMENT = True
 
 
-class SmartTurnParams(BaseModel):
+class SmartTurnParams(BaseTurnParams):
     """Configuration parameters for smart turn analysis.
 
     Parameters:
@@ -77,6 +79,9 @@ class BaseSmartTurn(BaseTurnAnalyzer):
         self._speech_triggered = False
         self._silence_ms = 0
         self._speech_start_time = 0
+        # Thread executor that will run the model. We only need one thread per
+        # analyzer because one analyzer just handles one audio stream.
+        self._executor = ThreadPoolExecutor(max_workers=1)
 
     @property
     def speech_triggered(self) -> bool:
@@ -151,7 +156,10 @@ class BaseSmartTurn(BaseTurnAnalyzer):
             Tuple containing the end-of-turn state and optional metrics data
             from the ML model analysis.
         """
-        state, result = await self._process_speech_segment(self._audio_buffer)
+        loop = asyncio.get_running_loop()
+        state, result = await loop.run_in_executor(
+            self._executor, self._process_speech_segment, self._audio_buffer
+        )
         if state == EndOfTurnState.COMPLETE or USE_ONLY_LAST_VAD_SEGMENT:
             self._clear(state)
         logger.debug(f"End of Turn result: {state}")
@@ -169,9 +177,7 @@ class BaseSmartTurn(BaseTurnAnalyzer):
         self._speech_start_time = 0
         self._silence_ms = 0
 
-    async def _process_speech_segment(
-        self, audio_buffer
-    ) -> Tuple[EndOfTurnState, Optional[MetricsData]]:
+    def _process_speech_segment(self, audio_buffer) -> Tuple[EndOfTurnState, Optional[MetricsData]]:
         """Process accumulated audio segment using ML model."""
         state = EndOfTurnState.INCOMPLETE
 
@@ -203,7 +209,7 @@ class BaseSmartTurn(BaseTurnAnalyzer):
         if len(segment_audio) > 0:
             start_time = time.perf_counter()
             try:
-                result = await self._predict_endpoint(segment_audio)
+                result = self._predict_endpoint(segment_audio)
                 state = (
                     EndOfTurnState.COMPLETE
                     if result["prediction"] == 1
@@ -249,6 +255,6 @@ class BaseSmartTurn(BaseTurnAnalyzer):
         return state, result_data
 
     @abstractmethod
-    async def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
+    def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
         """Predict end-of-turn using ML model from audio data."""
         pass

pipecat/audio/turn/smart_turn/http_smart_turn.py

@@ -104,11 +104,15 @@ class HttpSmartTurnAnalyzer(BaseSmartTurn):
             logger.error(f"Failed to send raw request to Daily Smart Turn: {e}")
             raise Exception("Failed to send raw request to Daily Smart Turn.")
 
-    async def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
+    def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
         """Predict end-of-turn using remote HTTP ML service."""
         try:
             serialized_array = self._serialize_array(audio_array)
-            return await self._send_raw_request(serialized_array)
+            loop = asyncio.get_running_loop()
+            future = asyncio.run_coroutine_threadsafe(
+                self._send_raw_request(serialized_array), loop
+            )
+            return future.result()
         except Exception as e:
             logger.error(f"Smart turn prediction failed: {str(e)}")
             # Return an incomplete prediction when a failure occurs

pipecat/audio/turn/smart_turn/local_smart_turn.py

@@ -64,7 +64,7 @@ class LocalSmartTurnAnalyzer(BaseSmartTurn):
         self._turn_model.eval()
         logger.debug("Loaded Local Smart Turn")
 
-    async def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
+    def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
         """Predict end-of-turn using local PyTorch model."""
         inputs = self._turn_processor(
             audio_array,

pipecat/audio/turn/smart_turn/local_smart_turn_v2.py

@@ -73,7 +73,7 @@ class LocalSmartTurnAnalyzerV2(BaseSmartTurn):
         self._turn_model.eval()
         logger.debug("Loaded Local Smart Turn v2")
 
-    async def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
+    def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
         """Predict end-of-turn using local PyTorch model."""
         inputs = self._turn_processor(
             audio_array,

pipecat/audio/turn/smart_turn/local_smart_turn_v3.py

@@ -0,0 +1,124 @@
+#
+# Copyright (c) 2025, Daily
+#
+# SPDX-License-Identifier: BSD 2-Clause License
+#
+
+"""Local turn analyzer for on-device ML inference using the smart-turn-v3 model.
+
+This module provides a smart turn analyzer that uses an ONNX model for
+local end-of-turn detection without requiring network connectivity.
+"""
+
+from typing import Any, Dict, Optional
+
+import numpy as np
+from loguru import logger
+
+from pipecat.audio.turn.smart_turn.base_smart_turn import BaseSmartTurn
+
+try:
+    import onnxruntime as ort
+    from transformers import WhisperFeatureExtractor
+except ModuleNotFoundError as e:
+    logger.error(f"Exception: {e}")
+    logger.error(
+        "In order to use LocalSmartTurnAnalyzerV3, you need to `pip install pipecat-ai[local-smart-turn-v3]`."
+    )
+    raise Exception(f"Missing module: {e}")
+
+
+class LocalSmartTurnAnalyzerV3(BaseSmartTurn):
+    """Local turn analyzer using the smart-turn-v3 ONNX model.
+
+    Provides end-of-turn detection using locally-stored ONNX model,
+    enabling offline operation without network dependencies.
+    """
+
+    def __init__(self, *, smart_turn_model_path: Optional[str] = None, **kwargs):
+        """Initialize the local ONNX smart-turn-v3 analyzer.
+
+        Args:
+            smart_turn_model_path: Path to the ONNX model file. If this is not
+                set, the bundled smart-turn-v3.0 model will be used.
+            **kwargs: Additional arguments passed to BaseSmartTurn.
+        """
+        super().__init__(**kwargs)
+
+        logger.debug("Loading Local Smart Turn v3 model...")
+
+        if not smart_turn_model_path:
+            # Load bundled model
+            model_name = "smart-turn-v3.0.onnx"
+            package_path = "pipecat.audio.turn.smart_turn.data"
+
+            try:
+                import importlib_resources as impresources
+
+                smart_turn_model_path = str(impresources.files(package_path).joinpath(model_name))
+            except BaseException:
+                from importlib import resources as impresources
+
+                try:
+                    with impresources.path(package_path, model_name) as f:
+                        smart_turn_model_path = f
+                except BaseException:
+                    smart_turn_model_path = str(
+                        impresources.files(package_path).joinpath(model_name)
+                    )
+
+        so = ort.SessionOptions()
+        so.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
+        so.inter_op_num_threads = 1
+        so.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+
+        self._feature_extractor = WhisperFeatureExtractor(chunk_length=8)
+        self._session = ort.InferenceSession(smart_turn_model_path, sess_options=so)
+
+        logger.debug("Loaded Local Smart Turn v3")
+
+    def _predict_endpoint(self, audio_array: np.ndarray) -> Dict[str, Any]:
+        """Predict end-of-turn using local ONNX model."""
+
+        def truncate_audio_to_last_n_seconds(audio_array, n_seconds=8, sample_rate=16000):
+            """Truncate audio to last n seconds or pad with zeros to meet n seconds."""
+            max_samples = n_seconds * sample_rate
+            if len(audio_array) > max_samples:
+                return audio_array[-max_samples:]
+            elif len(audio_array) < max_samples:
+                # Pad with zeros at the beginning
+                padding = max_samples - len(audio_array)
+                return np.pad(audio_array, (padding, 0), mode="constant", constant_values=0)
+            return audio_array
+
+        # Truncate to 8 seconds (keeping the end) or pad to 8 seconds
+        audio_array = truncate_audio_to_last_n_seconds(audio_array, n_seconds=8)
+
+        # Process audio using Whisper's feature extractor
+        inputs = self._feature_extractor(
+            audio_array,
+            sampling_rate=16000,
+            return_tensors="np",
+            padding="max_length",
+            max_length=8 * 16000,
+            truncation=True,
+            do_normalize=True,
+        )
+
+        # Extract features and ensure correct shape for ONNX
+        input_features = inputs.input_features.squeeze(0).astype(np.float32)
+        input_features = np.expand_dims(input_features, axis=0)  # Add batch dimension
+
+        # Run ONNX inference
+        outputs = self._session.run(None, {"input_features": input_features})
+
+        # Extract probability (ONNX model returns sigmoid probabilities)
+        probability = outputs[0][0].item()
+
+        # Make prediction (1 for Complete, 0 for Incomplete)
+        prediction = 1 if probability > 0.5 else 0
+
+        return {
+            "prediction": prediction,
+            "probability": probability,
+        }

pipecat/services/salesforce/__init__.py

@@ -0,0 +1,9 @@
+#
+# Copyright (c) 2024–2025, Daily
+#
+# SPDX-License-Identifier: BSD 2-Clause License
+#
+
+from .llm import SalesforceAgentLLMService
+
+__all__ = ["SalesforceAgentLLMService"]

pipecat/services/salesforce/llm.py

@@ -0,0 +1,587 @@
+#
+# Copyright (c) 2024–2025, Daily
+#
+# SPDX-License-Identifier: BSD 2-Clause License
+#
+
+"""Salesforce Agent API LLM service implementation."""
+
+import asyncio
+import json
+import time
+from dataclasses import dataclass
+from typing import AsyncGenerator, Dict, Optional
+
+import httpx
+from loguru import logger
+
+from pipecat.frames.frames import (
+    Frame,
+    LLMFullResponseEndFrame,
+    LLMFullResponseStartFrame,
+    LLMMessagesFrame,
+    LLMTextFrame,
+    LLMUpdateSettingsFrame,
+)
+from pipecat.processors.aggregators.openai_llm_context import (
+    OpenAILLMContext,
+    OpenAILLMContextFrame,
+)
+from pipecat.processors.frame_processor import FrameDirection
+from pipecat.services.llm_service import LLMService
+from pipecat.services.openai.llm import (
+    OpenAIAssistantContextAggregator,
+    OpenAIContextAggregatorPair,
+    OpenAIUserContextAggregator,
+)
+from pipecat.processors.aggregators.llm_response import (
+    LLMAssistantAggregatorParams,
+    LLMUserAggregatorParams,
+)
+from env_config import api_config
+from pipecat.utils.redis import create_async_redis_client
+
+
+@dataclass
+class SalesforceSessionInfo:
+    """Information about an active Salesforce Agent session."""
+
+    session_id: str
+    agent_id: str
+    created_at: float
+    last_used: float
+
+
+class SalesforceAgentLLMService(LLMService):
+    """Salesforce Agent API LLM service implementation.
+
+    This service integrates with Salesforce Agent API to provide conversational
+    AI capabilities using Salesforce's Agentforce platform.
+    """
+
+    def __init__(
+        self,
+        *,
+        model: str = "salesforce-agent",
+        session_timeout_secs: float = 3600.0,
+        agent_id: str = api_config.SALESFORCE_AGENT_ID,
+        org_domain: str = api_config.SALESFORCE_ORG_DOMAIN,
+        client_id: str = api_config.SALESFORCE_CLIENT_ID,
+        client_secret: str = api_config.SALESFORCE_CLIENT_SECRET,
+        api_host: str = api_config.SALESFORCE_API_HOST,
+        redis_url: Optional[str] = None,
+        **kwargs,
+    ):
+        """Initialize Salesforce Agent LLM service.
+
+        Reads configuration from environment variables:
+        - SALESFORCE_AGENT_ID: The Salesforce agent ID to interact with
+        - SALESFORCE_ORG_DOMAIN: Salesforce org domain (e.g., https://myorg.my.salesforce.com)
+        - SALESFORCE_CLIENT_ID: Connected app client ID for OAuth
+        - SALESFORCE_CLIENT_SECRET: Connected app client secret for OAuth
+        - SALESFORCE_API_HOST: Salesforce API host base URL (e.g., https://api.salesforce.com)
+
+        Args:
+            model: The model name (defaults to "salesforce-agent").
+            session_timeout_secs: Session timeout in seconds (default: 1 hour).
+            agent_id: Salesforce agent ID. Defaults to SALESFORCE_AGENT_ID.
+            org_domain: Salesforce org domain. Defaults to SALESFORCE_ORG_DOMAIN.
+            client_id: Salesforce connected app client ID. Defaults to SALESFORCE_CLIENT_ID.
+            client_secret: Salesforce connected app client secret. Defaults to SALESFORCE_CLIENT_SECRET.
+            api_host: Salesforce API host base URL. Defaults to SALESFORCE_API_HOST.
+            redis_url: Optional Redis URL override for token caching.
+            **kwargs: Additional arguments passed to parent LLMService.
+        """
+        # Initialize parent LLM service
+        super().__init__(**kwargs)
+
+        self._agent_id = agent_id
+        self._org_domain = org_domain 
+        self._client_id = client_id
+        self._client_secret = client_secret
+        self._api_host = api_host
+
+
+        # Validate required environment variables
+        required_vars = {
+            "SALESFORCE_AGENT_ID": self._agent_id,
+            "SALESFORCE_ORG_DOMAIN": self._org_domain,
+            "SALESFORCE_API_HOST": self._api_host,
+            "SALESFORCE_CLIENT_ID": self._client_id,
+            "SALESFORCE_CLIENT_SECRET": self._client_secret,
+        }
+
+        missing_vars = [var for var, value in required_vars.items() if not value]
+        if missing_vars:
+            raise ValueError(f"Missing required environment variables: {', '.join(missing_vars)}")
+
+        logger.info(f"Salesforce LLM initialized - Agent ID: {self._agent_id}")
+
+        self._session_timeout_secs = session_timeout_secs
+
+        if redis_url is not None:
+            self._redis_url = redis_url
+        else:
+            self._redis_url = getattr(api_config, "REDIS_URL", None)
+        self._redis_client = None
+        self._redis_client_init_attempted = False
+        self._token_cache_key = f"salesforce_agent_access_token:{self._agent_id}"
+        self._token_cache_leeway_secs = 300
+        self._sequence_counter = 0
+        self._warmup_task: Optional[asyncio.Task] = None
+
+        # Session management
+        self._sessions: Dict[str, SalesforceSessionInfo] = {}
+        self._current_session_id: Optional[str] = None
+
+        # HTTP client for API calls
+        self._http_client = httpx.AsyncClient(
+            timeout=30.0,
+            limits=httpx.Limits(
+                max_keepalive_connections=10,
+                max_connections=100,
+                keepalive_expiry=None,
+            ),
+        )
+
+        self._schedule_session_warmup()
+        
+
+    async def __aenter__(self):
+        """Async context manager entry."""
+        await self.ensure_session_ready()
+        return self
+
+    async def __aexit__(self, exc_type, exc_val, exc_tb):
+        """Async context manager exit."""
+        if self._warmup_task:
+            try:
+                await asyncio.shield(self._warmup_task)
+            except Exception as exc:  # pragma: no cover - warmup best effort
+                logger.debug(f"Salesforce warmup task failed during exit: {exc}")
+            finally:
+                self._warmup_task = None
+
+        await self._cleanup_sessions()
+        await self._http_client.aclose()
+
+        if self._redis_client:
+            close_coro = getattr(self._redis_client, "close", None)
+            if callable(close_coro):
+                try:
+                    await close_coro()
+                except Exception as exc:  # pragma: no cover - best effort cleanup
+                    logger.debug(f"Failed to close Redis client cleanly: {exc}")
+        self._redis_client = None
+        self._redis_client_init_attempted = False
+
+    def can_generate_metrics(self) -> bool:
+        """Check if this service can generate processing metrics."""
+        return True
+
+    def _schedule_session_warmup(self):
+        """Kick off background warm-up if an event loop is running."""
+        try:
+            loop = asyncio.get_running_loop()
+        except RuntimeError:
+            return
+
+        if loop.is_closed():
+            return
+
+        async def _warmup():
+            try:
+                await self.ensure_session_ready()
+            except Exception as exc:  # pragma: no cover - warmup best effort
+                logger.warning(f"Salesforce warmup failed: {exc}")
+                raise
+
+        task = loop.create_task(_warmup())
+
+        def _on_done(warmup_task: asyncio.Task):
+            if warmup_task.cancelled():
+                logger.debug("Salesforce warmup task cancelled")
+            elif warmup_task.exception():
+                logger.warning(f"Salesforce warmup task error: {warmup_task.exception()}")
+            self._warmup_task = None
+
+        task.add_done_callback(_on_done)
+        self._warmup_task = task
+
+    def _get_redis_client(self):
+        """Return a Redis client for token caching if configured."""
+        if self._redis_client is None and not self._redis_client_init_attempted:
+            self._redis_client_init_attempted = True
+            self._redis_client = create_async_redis_client(
+                self._redis_url, decode_responses=True, encoding="utf-8", logger=logger
+            )
+
+        return self._redis_client
+
+    async def _get_cached_access_token(self) -> Optional[str]:
+        """Return cached access token from Redis."""
+        redis_client = self._get_redis_client()
+        if not redis_client:
+            return None
+
+        try:
+            return await redis_client.get(self._token_cache_key)
+        except Exception as exc:  # pragma: no cover - cache failures shouldn't break flow
+            logger.warning(f"Failed to read Salesforce token from Redis: {exc}")
+            return None
+
+    async def _set_cached_access_token(self, token: str, expires_in: Optional[int]):
+        """Persist access token in Redis with TTL matching Salesforce expiry."""
+        redis_client = self._get_redis_client()
+        if not redis_client:
+            return
+
+        ttl_seconds = 3600
+        if expires_in is not None:
+            try:
+                ttl_seconds = max(int(expires_in) - self._token_cache_leeway_secs, 30)
+            except (TypeError, ValueError):
+                logger.debug("Unable to parse Salesforce token expiry; falling back to default TTL")
+
+        try:
+            await redis_client.set(self._token_cache_key, token, ex=ttl_seconds)
+        except Exception as exc:  # pragma: no cover - cache failures shouldn't break flow
+            logger.warning(f"Failed to store Salesforce token in Redis: {exc}")
+
+    async def _get_access_token(self) -> str:
+        """Get OAuth access token using client credentials."""
+        cached_token = await self._get_cached_access_token()
+        if cached_token:
+            return cached_token
+
+        token_url = f"{self._org_domain}/services/oauth2/token"
+        data = {
+            "grant_type": "client_credentials",
+            "client_id": self._client_id,
+            "client_secret": self._client_secret,
+        }
+
+        try:
+            response = await self._http_client.post(token_url, data=data)
+            response.raise_for_status()
+            token_data = response.json()
+            access_token = token_data["access_token"]
+            await self._set_cached_access_token(access_token, token_data.get("expires_in"))
+            return access_token
+        except Exception as e:
+            logger.error(f"Failed to get access token: {e}")
+            raise
+
+    async def _create_session(self) -> str:
+        """Create a new Salesforce Agent session."""
+        access_token = await self._get_access_token()
+        session_url = f"{self._api_host}/einstein/ai-agent/v1/agents/{self._agent_id}/sessions"
+
+        headers = {
+            "Authorization": f"Bearer {access_token}",
+            "Content-Type": "application/json",
+        }
+
+        external_session_key = f"pipecat-{int(time.time())}-{id(self)}"
+
+        payload = {
+            "externalSessionKey": external_session_key,
+            "instanceConfig": {"endpoint": self._org_domain},
+            "tz": "America/Los_Angeles",
+            "variables": [{"name": "$Context.EndUserLanguage", "type": "Text", "value": "en_US"}],
+            "featureSupport": "Streaming",
+            "streamingCapabilities": {"chunkTypes": ["Text"]},
+            "bypassUser": True,
+        }
+
+        try:
+            response = await self._http_client.post(session_url, headers=headers, json=payload)
+            response.raise_for_status()
+            session_data = response.json()
+            session_id = session_data["sessionId"]
+
+            # Store session info
+            current_time = time.time()
+            self._sessions[session_id] = SalesforceSessionInfo(
+                session_id=session_id,
+                agent_id=self._agent_id,
+                created_at=current_time,
+                last_used=current_time,
+            )
+
+            logger.debug(f"Created Salesforce Agent session: {session_id}")
+            return session_id
+
+        except Exception as e:
+            logger.error(f"Failed to create Salesforce Agent session: {e}")
+            raise
+
+    async def _get_or_create_session(self) -> str:
+        """Get existing session or create a new one."""
+        current_time = time.time()
+
+        # Check if current session is still valid
+        if self._current_session_id and self._current_session_id in self._sessions:
+            session = self._sessions[self._current_session_id]
+            if current_time - session.last_used < self._session_timeout_secs:
+                session.last_used = current_time
+                return self._current_session_id
+            else:
+                # Session expired, remove it
+                self._sessions.pop(self._current_session_id, None)
+                self._current_session_id = None
+
+        # Create new session
+        self._current_session_id = await self._create_session()
+        return self._current_session_id
+
+    async def ensure_session_ready(self) -> str:
+        """Ensure a Salesforce session is ready for use."""
+        return await self._get_or_create_session()
+
+    async def _cleanup_sessions(self):
+        """Clean up expired sessions."""
+        current_time = time.time()
+        expired_sessions = []
+
+        for session_id, session in self._sessions.items():
+            if current_time - session.last_used > self._session_timeout_secs:
+                expired_sessions.append(session_id)
+
+        for session_id in expired_sessions:
+            try:
+                # End the session via API
+                access_token = await self._get_access_token()
+                url = f"{self._api_host}/einstein/ai-agent/v1/sessions/{session_id}"
+                headers = {
+                    "Authorization": f"Bearer {access_token}",
+                    "x-session-end-reason": "UserRequest",
+                }
+                await self._http_client.delete(url, headers=headers)
+            except Exception as e:
+                logger.warning(f"Failed to end session {session_id}: {e}")
+            finally:
+                self._sessions.pop(session_id, None)
+                if self._current_session_id == session_id:
+                    self._current_session_id = None
+
+    def _extract_user_message(self, context: OpenAILLMContext) -> str:
+        """Extract the last user message from context.
+        
+        Similar to Vistaar pattern - extract only the most recent user message.
+        
+        Args:
+            context: The OpenAI LLM context containing messages.
+            
+        Returns:
+            The last user message as a string.
+        """
+        messages = context.get_messages()
+        
+        # Find the last user message (iterate in reverse for efficiency)
+        for message in reversed(messages):
+            if message.get("role") == "user":
+                content = message.get("content", "")
+                
+                # Handle content that might be a list (for multimodal messages)
+                if isinstance(content, list):
+                    text_parts = [
+                        item.get("text", "") for item in content if item.get("type") == "text"
+                    ]
+                    content = " ".join(text_parts)
+                
+                if isinstance(content, str):
+                    return content.strip()
+        
+        return ""
+
+    def _generate_sequence_id(self) -> int:
+        """Generate a sequence ID for the message."""
+        self._sequence_counter += 1
+        return self._sequence_counter
+
+    async def _stream_salesforce_response(self, session_id: str, user_message: str) -> AsyncGenerator[str, None]:
+        """Stream response from Salesforce Agent API."""
+        access_token = await self._get_access_token()
+        url = f"{self._api_host}/einstein/ai-agent/v1/sessions/{session_id}/messages/stream"
+        
+        headers = {
+            "Authorization": f"Bearer {access_token}",
+            "Content-Type": "application/json",
+            "Accept": "text/event-stream",
+        }
+
+        message_data = {
+            "message": {
+                "sequenceId": self._generate_sequence_id(),
+                "type": "Text",
+                "text": user_message
+            },
+            "variables": [
+                {
+                    "name": "$Context.EndUserLanguage",
+                    "type": "Text",
+                    "value": "en_US"
+                }
+            ]
+        }
+
+        try:
+            logger.info(f"🌐 Salesforce API request: {user_message[:50]}...")
+            async with self._http_client.stream("POST", url, headers=headers, json=message_data) as response:
+                response.raise_for_status()
+                
+                async for line in response.aiter_lines():
+                    if not line:
+                        continue
+
+                    # Parse SSE format
+                    if line.startswith("data: "):
+                        try:
+                            data = json.loads(line[6:])
+                            message = data.get("message", {})
+                            message_type = message.get("type")
+
+                            if message_type == "TextChunk":
+                                content = message.get("text", "") or message.get("message", "")
+                                if content:
+                                    yield content
+                            elif message_type == "EndOfTurn":
+                                logger.info("🏁 Salesforce response complete")
+                                break
+                            elif message_type == "Inform":
+                                # Skip INFORM events to avoid duplication
+                                continue
+
+                        except json.JSONDecodeError as e:
+                            logger.warning(f"JSON decode error: {e}, line: {line}")
+                            continue
+
+        except Exception as e:
+            logger.error(f"Failed to stream from Salesforce Agent API: {e}")
+            raise
+
+    async def _process_context(self, context: OpenAILLMContext):
+        """Process the LLM context and generate streaming response.
+
+        Args:
+            context: The OpenAI LLM context containing messages to process.
+        """
+        logger.info(f"🔄 Salesforce processing context with {len(context.get_messages())} messages")
+        
+        # Extract user message from context first
+        user_message = self._extract_user_message(context)
+        
+        if not user_message:
+            logger.warning("Salesforce: No user message found in context")
+            return
+        
+        try:
+            logger.info(f"🎯 Salesforce extracted query: {user_message}")
+            
+            # Start response 
+            await self.push_frame(LLMFullResponseStartFrame())
+            await self.push_frame(LLMFullResponseStartFrame(),FrameDirection.UPSTREAM)
+            await self.start_processing_metrics()
+            await self.start_ttfb_metrics()
+            
+            # Get or create session
+            session_id = await self._get_or_create_session()
+            
+            first_chunk = True
+            
+            # Stream the response
+            async for text_chunk in self._stream_salesforce_response(session_id, user_message):
+                if first_chunk:
+                    await self.stop_ttfb_metrics()
+                    first_chunk = False
+                
+                # Push each text chunk as it arrives
+                await self.push_frame(LLMTextFrame(text=text_chunk))
+            
+        except Exception as e:
+            logger.error(f"Salesforce context processing error: {type(e).__name__}: {str(e)}")
+            import traceback
+            logger.error(f"Salesforce traceback: {traceback.format_exc()}")
+            raise
+        finally:
+            await self.stop_processing_metrics()
+            await self.push_frame(LLMFullResponseEndFrame())
+            await self.push_frame(LLMFullResponseEndFrame(), FrameDirection.UPSTREAM)
+
+    async def process_frame(self, frame: Frame, direction: FrameDirection):
+        """Process frames for LLM completion requests.
+
+        Args:
+            frame: The frame to process.
+            direction: The direction of frame processing.
+        """
+        context = None
+        if isinstance(frame, OpenAILLMContextFrame):
+            context = frame.context
+            logger.info(f"🔍 Received OpenAILLMContextFrame with {len(context.get_messages())} messages")
+        elif isinstance(frame, LLMMessagesFrame):
+            context = OpenAILLMContext.from_messages(frame.messages)
+            logger.info(f"🔍 Received LLMMessagesFrame with {len(frame.messages)} messages")
+        elif isinstance(frame, LLMUpdateSettingsFrame):
+            # Call super for settings frames and update settings
+            await super().process_frame(frame, direction)
+            settings = frame.settings
+            logger.debug(f"Updated Salesforce settings: {settings}")
+        else:
+            # For non-context frames, call super and push them downstream
+            await super().process_frame(frame, direction)
+            await self.push_frame(frame, direction)
+
+        if context:
+            try:
+                await self._process_context(context)
+            except httpx.TimeoutException:
+                logger.error("Timeout while processing Salesforce request")
+                await self._call_event_handler("on_completion_timeout")
+            except Exception as e:
+                logger.error(f"Error processing Salesforce request: {e}")
+                raise
+
+    def create_context_aggregator(
+        self,
+        context: OpenAILLMContext,
+        *,
+        user_params: LLMUserAggregatorParams = LLMUserAggregatorParams(),
+        assistant_params: LLMAssistantAggregatorParams = LLMAssistantAggregatorParams(),
+    ) -> OpenAIContextAggregatorPair:
+        """Create context aggregators for Salesforce LLM.
+
+        Since Salesforce uses OpenAI-compatible message format, we reuse OpenAI's
+        context aggregators directly
+
+        Args:
+            context: The LLM context to create aggregators for.
+            user_params: Parameters for user message aggregation.
+            assistant_params: Parameters for assistant message aggregation.
+
+        Returns:
+            OpenAIContextAggregatorPair: A pair of OpenAI context aggregators,
+            compatible with Salesforce's OpenAI-like message format.
+        """
+        context.set_llm_adapter(self.get_llm_adapter())
+        user = OpenAIUserContextAggregator(context, params=user_params)
+        assistant = OpenAIAssistantContextAggregator(context, params=assistant_params)
+        return OpenAIContextAggregatorPair(_user=user, _assistant=assistant)
+
+    def get_llm_adapter(self):
+        """Get the LLM adapter for this service."""
+        from pipecat.adapters.services.open_ai_adapter import OpenAILLMAdapter
+        return OpenAILLMAdapter()
+
+    async def close(self):
+        """Close the HTTP client when the service is destroyed."""
+        await self._cleanup_sessions()
+        await self._http_client.aclose()
+
+    def __del__(self):
+        """Ensure the client is closed on deletion."""
+        try:
+            asyncio.create_task(self._http_client.aclose())
+        except:
+            pass

pipecat/utils/redis.py

@@ -0,0 +1,58 @@
+"""Async Redis helper utilities."""
+
+from __future__ import annotations
+
+from typing import Any, Optional, TYPE_CHECKING
+
+from urllib.parse import urlparse
+
+try:
+    import redis.asyncio as redis
+except ImportError:  # pragma: no cover - Redis is optional
+    redis = None
+
+
+if TYPE_CHECKING:  # pragma: no cover - typing aid
+    from redis.asyncio import Redis
+
+
+def create_async_redis_client(
+    url: Optional[str],
+    *,
+    decode_responses: bool = True,
+    encoding: str = "utf-8",
+    logger: Optional[Any] = None,
+    **kwargs,
+) -> Optional["Redis"]:
+    """Return a configured async Redis client or None if unavailable.
+
+    Args:
+        url: Redis connection URL.
+        decode_responses: Whether to decode responses to str.
+        encoding: Character encoding to use with decoded responses.
+        logger: Optional logger supporting .warning() for diagnostics.
+        **kwargs: Additional keyword arguments forwarded to Redis.from_url.
+    """
+    if redis is None:
+        return None
+
+    if not url or url in {"redis_url", "REDIS_URL"}:
+        return None
+
+    parsed = urlparse(url)
+    connection_kwargs = {
+        "decode_responses": decode_responses,
+        "encoding": encoding,
+    }
+    connection_kwargs.update(kwargs)
+
+    if parsed.scheme == "rediss":
+        connection_kwargs.setdefault("ssl_cert_reqs", "none")
+        connection_kwargs.setdefault("ssl_check_hostname", False)
+
+    try:
+        return redis.Redis.from_url(url, **connection_kwargs)
+    except Exception as exc:  # pragma: no cover - best effort logging
+        if logger is not None:
+            logger.warning(f"Failed to create Redis client: {exc}")
+        return None