Spaces:
Running
Running
File size: 14,602 Bytes
53ea588 e9446cb 53ea588 e9446cb 53ea588 e9446cb 53ea588 e9446cb 53ea588 e9446cb 53ea588 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 |
"""LangGraph-backed LLM service for Pipecat pipelines.
This service adapts a running LangGraph agent (accessed via langgraph-sdk)
to Pipecat's frame-based processing model. It consumes `OpenAILLMContextFrame`
or `LLMMessagesFrame` inputs, extracts the latest user message (using the
LangGraph server's thread to persist history), and streams assistant tokens
back as `LLMTextFrame` until completion.
"""
from __future__ import annotations
import asyncio
from typing import Any, Optional
import os
from dotenv import load_dotenv
from langgraph_sdk import get_client
from langchain_core.messages import HumanMessage
from loguru import logger
from pipecat.frames.frames import (
Frame,
LLMFullResponseEndFrame,
LLMFullResponseStartFrame,
LLMMessagesFrame,
LLMTextFrame,
StartInterruptionFrame,
VisionImageRawFrame,
)
from pipecat.processors.aggregators.openai_llm_context import OpenAILLMContext, OpenAILLMContextFrame
from pipecat.processors.frame_processor import FrameDirection
from pipecat.services.openai.llm import OpenAILLMService
load_dotenv()
# TTS sanitize helper: normalize curly quotes/dashes and non-breaking spaces to ASCII
def _tts_sanitize(text: str) -> str:
try:
if not isinstance(text, str):
text = str(text)
replacements = {
"\u2018": "'", # left single quote
"\u2019": "'", # right single quote / apostrophe
"\u201C": '"', # left double quote
"\u201D": '"', # right double quote
"\u00AB": '"', # left angle quote
"\u00BB": '"', # right angle quote
"\u2013": "-", # en dash
"\u2014": "-", # em dash
"\u2026": "...",# ellipsis
"\u00A0": " ", # non-breaking space
"\u202F": " ", # narrow no-break space
}
for k, v in replacements.items():
text = text.replace(k, v)
return text
except Exception:
return text
class LangGraphLLMService(OpenAILLMService):
"""Pipecat LLM service that delegates responses to a LangGraph agent.
Attributes:
base_url: LangGraph API base URL, e.g. "http://127.0.0.1:2024".
assistant: Assistant name or id registered with the LangGraph server.
user_email: Value for `configurable.user_email` (routing / personalization).
stream_mode: SDK stream mode ("updates", "values", "messages", "events").
debug_stream: When True, logs raw stream events for troubleshooting.
"""
def __init__(
self,
*,
base_url: str = "http://127.0.0.1:2024",
assistant: str = "ace-base-agent",
user_email: str = "test@example.com",
stream_mode: str = "values",
debug_stream: bool = False,
thread_id: Optional[str] = None,
auth_token: Optional[str] = None,
**kwargs: Any,
) -> None:
# Initialize base class; OpenAI settings unused but required by parent
super().__init__(api_key="", **kwargs)
self.base_url = base_url
self.assistant = assistant
self.user_email = user_email
self.stream_mode = stream_mode
self.debug_stream = debug_stream
# Optional auth header
token = (
auth_token
or os.getenv("LANGGRAPH_AUTH_TOKEN")
or os.getenv("AUTH0_ACCESS_TOKEN")
or os.getenv("AUTH_BEARER_TOKEN")
)
headers = {"Authorization": f"Bearer {token}"} if isinstance(token, str) and token else None
self._client = get_client(url=self.base_url, headers=headers) if headers else get_client(url=self.base_url)
self._thread_id: Optional[str] = thread_id
self._current_task: Optional[asyncio.Task] = None
self._outer_open: bool = False
self._emitted_texts: set[str] = set()
async def _ensure_thread(self) -> Optional[str]:
if self._thread_id:
return self._thread_id
try:
thread = await self._client.threads.create()
except Exception as exc: # noqa: BLE001
logger.warning(f"LangGraph: failed to create thread; proceeding threadless. Error: {exc}")
self._thread_id = None
return None
thread_id = getattr(thread, "thread_id", None)
if thread_id is None and isinstance(thread, dict):
thread_id = thread.get("thread_id") or thread.get("id")
if thread_id is None:
thread_id = getattr(thread, "id", None)
if isinstance(thread_id, str) and thread_id:
self._thread_id = thread_id
else:
logger.warning("LangGraph: could not determine thread id; proceeding threadless.")
self._thread_id = None
return self._thread_id
@staticmethod
def _extract_latest_user_text(context: OpenAILLMContext) -> str:
"""Return the latest user (or fallback system) message content.
The LangGraph server maintains history via threads, so we only need to
send the current turn text. Prefer the latest user message; if absent,
fall back to the latest system message so system-only kickoffs can work.
"""
messages = context.get_messages() or []
for msg in reversed(messages):
try:
if msg.get("role") == "user":
content = msg.get("content", "")
return content if isinstance(content, str) else str(content)
except Exception: # Defensive against unexpected shapes
continue
# Fallback: use the most recent system message if no user message exists
for msg in reversed(messages):
try:
if msg.get("role") == "system":
content = msg.get("content", "")
return content if isinstance(content, str) else str(content)
except Exception:
continue
return ""
async def _stream_langgraph(self, text: str) -> None:
config = {"configurable": {"user_email": self.user_email}}
# Attempt to ensure thread; OK if None (threadless run)
await self._ensure_thread()
try:
async for chunk in self._client.runs.stream(
self._thread_id,
self.assistant,
input=[HumanMessage(content=text)],
stream_mode=self.stream_mode,
config=config,
):
data = getattr(chunk, "data", None)
event = getattr(chunk, "event", "") or ""
if self.debug_stream:
try:
# Short, structured debugging output
dtype = type(data).__name__
preview = ""
if hasattr(data, "content") and isinstance(getattr(data, "content"), str):
c = getattr(data, "content")
preview = c[:120]
elif isinstance(data, dict):
preview = ",".join(list(data.keys())[:6])
logger.debug(f"[LangGraph stream] event={event} data={dtype}:{preview}")
except Exception: # noqa: BLE001
logger.debug(f"[LangGraph stream] event={event}")
# Token streaming events (LangChain chat model streaming)
if "on_chat_model_stream" in event or event.endswith(".on_chat_model_stream"):
part_text = ""
d = data
if isinstance(d, dict):
if "chunk" in d:
ch = d["chunk"]
part_text = getattr(ch, "content", None) or ""
if not isinstance(part_text, str):
part_text = str(part_text)
elif "delta" in d:
delta = d["delta"]
part_text = getattr(delta, "content", None) or ""
if not isinstance(part_text, str):
part_text = str(part_text)
elif "content" in d and isinstance(d["content"], str):
part_text = d["content"]
else:
part_text = getattr(d, "content", "")
if part_text:
if not self._outer_open:
await self.push_frame(LLMFullResponseStartFrame())
self._outer_open = True
self._emitted_texts.clear()
if part_text not in self._emitted_texts:
self._emitted_texts.add(part_text)
await self.push_frame(LLMTextFrame(_tts_sanitize(part_text)))
# Final value-style events (values mode)
if event == "values":
# Some dev servers send final AI message content here
final_text = ""
if hasattr(data, "content") and isinstance(getattr(data, "content"), str):
final_text = getattr(data, "content")
elif isinstance(data, dict):
c = data.get("content")
if isinstance(c, str):
final_text = c
if final_text:
# Close backchannel utterance if open
if self._outer_open:
await self.push_frame(LLMFullResponseEndFrame())
self._outer_open = False
self._emitted_texts.clear()
# Emit final explanation as its own message
await self.push_frame(LLMFullResponseStartFrame())
await self.push_frame(LLMTextFrame(_tts_sanitize(final_text)))
await self.push_frame(LLMFullResponseEndFrame())
# Messages mode: look for an array of messages
if event == "messages" or event.endswith(":messages"):
try:
msgs = None
if isinstance(data, dict):
msgs = data.get("messages") or data.get("result") or data.get("value")
elif hasattr(data, "messages"):
msgs = getattr(data, "messages")
if isinstance(msgs, list) and msgs:
last = msgs[-1]
content = getattr(last, "content", None)
if content is None and isinstance(last, dict):
content = last.get("content")
if isinstance(content, str) and content:
if not self._outer_open:
await self.push_frame(LLMFullResponseStartFrame())
self._outer_open = True
self._emitted_texts.clear()
if content not in self._emitted_texts:
self._emitted_texts.add(content)
await self.push_frame(LLMTextFrame(_tts_sanitize(content)))
except Exception as exc: # noqa: BLE001
logger.debug(f"LangGraph messages parsing error: {exc}")
# If payload is a plain string, emit it
if isinstance(data, str):
txt = data.strip()
if txt:
if not self._outer_open:
await self.push_frame(LLMFullResponseStartFrame())
self._outer_open = True
self._emitted_texts.clear()
if txt not in self._emitted_texts:
self._emitted_texts.add(txt)
await self.push_frame(LLMTextFrame(_tts_sanitize(txt)))
except Exception as exc: # noqa: BLE001
logger.error(f"LangGraph stream error: {exc}")
async def _process_context_and_frames(self, context: OpenAILLMContext) -> None:
"""Adapter entrypoint: push start/end frames and stream tokens."""
try:
# Defer opening until backchannels arrive; final will be emitted separately
user_text = self._extract_latest_user_text(context)
if not user_text:
logger.debug("LangGraph: no user text in context; skipping run.")
return
self._outer_open = False
self._emitted_texts.clear()
await self._stream_langgraph(user_text)
finally:
if self._outer_open:
await self.push_frame(LLMFullResponseEndFrame())
self._outer_open = False
async def process_frame(self, frame: Frame, direction: FrameDirection):
"""Process pipeline frames, handling interruptions and context inputs."""
context: Optional[OpenAILLMContext] = None
if isinstance(frame, OpenAILLMContextFrame):
context = frame.context
elif isinstance(frame, LLMMessagesFrame):
context = OpenAILLMContext.from_messages(frame.messages)
elif isinstance(frame, VisionImageRawFrame):
# Not implemented for LangGraph adapter; ignore images
context = None
elif isinstance(frame, StartInterruptionFrame):
# Relay interruption downstream and cancel any active run
await self._start_interruption()
await self.stop_all_metrics()
await self.push_frame(frame, direction)
if self._current_task is not None and not self._current_task.done():
await self.cancel_task(self._current_task)
self._current_task = None
return
else:
await super().process_frame(frame, direction)
if context is not None:
if self._current_task is not None and not self._current_task.done():
await self.cancel_task(self._current_task)
self._current_task = None
logger.debug("LangGraph LLM: canceled previous task")
self._current_task = self.create_task(self._process_context_and_frames(context))
self._current_task.add_done_callback(lambda _: setattr(self, "_current_task", None))
|