99 lines
3.9 KiB
Python
99 lines
3.9 KiB
Python
"""Canonical serialization for LangChain / LangGraph objects.
|
|
|
|
Provides a single source of truth for converting LangChain message
|
|
objects, Pydantic models, and LangGraph state dicts into plain
|
|
JSON-serialisable Python structures.
|
|
|
|
Consumers: ``deerflow.runtime.runs.worker`` (SSE publishing) and
|
|
``app.gateway.routers.threads`` (REST responses).
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import dataclasses
|
|
from typing import Any
|
|
|
|
|
|
def serialize_lc_object(obj: Any) -> Any:
|
|
"""Recursively serialize a LangChain object to a JSON-serialisable dict."""
|
|
if obj is None:
|
|
return None
|
|
if isinstance(obj, (str, int, float, bool)):
|
|
return obj
|
|
if isinstance(obj, dict):
|
|
return {k: serialize_lc_object(v) for k, v in obj.items()}
|
|
if isinstance(obj, (list, tuple)):
|
|
return [serialize_lc_object(item) for item in obj]
|
|
# Pydantic v2
|
|
if hasattr(obj, "model_dump"):
|
|
try:
|
|
return obj.model_dump()
|
|
except Exception:
|
|
pass
|
|
# Pydantic v1 / older objects
|
|
if hasattr(obj, "dict"):
|
|
try:
|
|
return obj.dict()
|
|
except Exception:
|
|
pass
|
|
# Dataclass instances —— LangGraph 的 ``Interrupt`` 就是 dataclass 没有
|
|
# model_dump 也没有 dict 方法；之前 fallback 到 ``str(obj)`` 会把
|
|
# ``Interrupt(value=..., id='...')`` 序列化成字符串，下游 SDK
|
|
# (``@langchain/langgraph-sdk``) 期待 dict 结构 → 直接丢弃 → AI 写作
|
|
# 的 4 个干预卡永远不弹（参 ai-writing-优化.md §阶段 3 排查记录）。
|
|
if dataclasses.is_dataclass(obj) and not isinstance(obj, type):
|
|
try:
|
|
return {k: serialize_lc_object(v) for k, v in dataclasses.asdict(obj).items()}
|
|
except Exception:
|
|
pass
|
|
# Last resort
|
|
try:
|
|
return str(obj)
|
|
except Exception:
|
|
return repr(obj)
|
|
|
|
|
|
def serialize_channel_values(channel_values: dict[str, Any]) -> dict[str, Any]:
|
|
"""Serialize channel values, stripping internal LangGraph keys.
|
|
|
|
``__pregel_*`` 键始终剥离。``__interrupt__`` 仅在**为空时**剥离 —— 因为
|
|
``@langchain/langgraph-sdk`` 的 ``useStream`` 只在 ``event === "values"``
|
|
通道里读 ``data.__interrupt__`` 并通过 ``stream.interrupt`` 暴露给上层。
|
|
AI 写作的 4 个干预卡（material/outline/draft/review）就是靠它触发的；
|
|
继续过滤会让前端永远拿不到 interrupt。
|
|
|
|
主聊天不消费 ``stream.interrupt``，但若我们让一个空数组溜过去，SDK 会
|
|
fallback 成 ``{when: "breakpoint"}`` 的伪 interrupt —— 所以仅在数组非空
|
|
时保留，两侧都不破坏。
|
|
"""
|
|
result: dict[str, Any] = {}
|
|
for key, value in channel_values.items():
|
|
if key.startswith("__pregel_"):
|
|
continue
|
|
if key == "__interrupt__" and not value:
|
|
continue
|
|
result[key] = serialize_lc_object(value)
|
|
return result
|
|
|
|
|
|
def serialize_messages_tuple(obj: Any) -> Any:
|
|
"""Serialize a messages-mode tuple ``(chunk, metadata)``."""
|
|
if isinstance(obj, tuple) and len(obj) == 2:
|
|
chunk, metadata = obj
|
|
return [serialize_lc_object(chunk), metadata if isinstance(metadata, dict) else {}]
|
|
return serialize_lc_object(obj)
|
|
|
|
|
|
def serialize(obj: Any, *, mode: str = "") -> Any:
|
|
"""Serialize LangChain objects with mode-specific handling.
|
|
|
|
* ``messages`` — obj is ``(message_chunk, metadata_dict)``
|
|
* ``values`` — obj is the full state dict; ``__pregel_*`` keys stripped
|
|
* everything else — recursive ``model_dump()`` / ``dict()`` fallback
|
|
"""
|
|
if mode == "messages":
|
|
return serialize_messages_tuple(obj)
|
|
if mode == "values":
|
|
return serialize_channel_values(obj) if isinstance(obj, dict) else serialize_lc_object(obj)
|
|
return serialize_lc_object(obj)
|