379 lines
13 KiB
Python

from __future__ import annotations
from collections.abc import AsyncIterator, Callable, Iterator, Sequence
from dataclasses import fields, is_dataclass
from typing import (
Any,
TypeVar,
cast,
)
from uuid import UUID, uuid4
from langchain_core.callbacks import BaseCallbackHandler
from langchain_core.messages import BaseMessage
from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, LLMResult
from pydantic import BaseModel
from langgraph._internal._constants import NS_SEP
from langgraph.constants import TAG_HIDDEN, TAG_NOSTREAM
from langgraph.pregel.protocol import StreamChunk
from langgraph.types import Command
try:
from langchain_core.tracers._streaming import _StreamingCallbackHandler
except ImportError:
_StreamingCallbackHandler = object # type: ignore
try:
from langchain_core.tracers._streaming import _V2StreamingCallbackHandler
except ImportError:
_V2StreamingCallbackHandler = object # type: ignore
T = TypeVar("T")
Meta = tuple[tuple[str, ...], dict[str, Any]]
def _state_values(obj: Any) -> Sequence[Any]:
"""Extract top-level field values from a state object (dict, BaseModel, or dataclass)."""
if isinstance(obj, dict):
return list(obj.values())
elif isinstance(obj, BaseModel):
return [getattr(obj, k) for k in type(obj).model_fields]
elif is_dataclass(obj) and not isinstance(obj, type):
return [getattr(obj, f.name) for f in fields(obj)]
return ()
class StreamMessagesHandler(BaseCallbackHandler, _StreamingCallbackHandler):
"""A callback handler that implements stream_mode=messages.
Collects messages from:
(1) chat model stream events; and
(2) node outputs.
"""
run_inline = True
"""We want this callback to run in the main thread to avoid order/locking issues."""
def __init__(
self,
stream: Callable[[StreamChunk], None],
subgraphs: bool,
*,
parent_ns: tuple[str, ...] | None = None,
) -> None:
"""Configure the handler to stream messages from LLMs and nodes.
Args:
stream: A callable that takes a StreamChunk and emits it.
subgraphs: Whether to emit messages from subgraphs.
parent_ns: The namespace where the handler was created.
We keep track of this namespace to allow calls to subgraphs that
were explicitly requested as a stream with `messages` mode
configured.
Example:
parent_ns is used to handle scenarios where the subgraph is explicitly
streamed with `stream_mode="messages"`.
```python
def parent_graph_node():
# This node is in the parent graph.
async for event in some_subgraph(..., stream_mode="messages"):
do something with event # <-- these events will be emitted
return ...
parent_graph.invoke(subgraphs=False)
```
"""
self.stream = stream
self.subgraphs = subgraphs
self.metadata: dict[UUID, Meta] = {}
self.seen: set[int | str] = set()
self.parent_ns = parent_ns
def _emit(self, meta: Meta, message: BaseMessage, *, dedupe: bool = False) -> None:
if dedupe and message.id in self.seen:
return
else:
if message.id is None:
message.id = str(uuid4())
self.seen.add(message.id)
self.stream((meta[0], "messages", (message, meta[1])))
def _find_and_emit_messages(self, meta: Meta, response: Any) -> None:
if isinstance(response, BaseMessage):
self._emit(meta, response, dedupe=True)
elif isinstance(response, Sequence):
for value in response:
if isinstance(value, BaseMessage):
self._emit(meta, value, dedupe=True)
else:
for value in _state_values(response):
if isinstance(value, BaseMessage):
self._emit(meta, value, dedupe=True)
elif isinstance(value, Sequence):
for item in value:
if isinstance(item, BaseMessage):
self._emit(meta, item, dedupe=True)
def tap_output_aiter(
self, run_id: UUID, output: AsyncIterator[T]
) -> AsyncIterator[T]:
return output
def tap_output_iter(self, run_id: UUID, output: Iterator[T]) -> Iterator[T]:
return output
def on_chat_model_start(
self,
serialized: dict[str, Any],
messages: list[list[BaseMessage]],
*,
run_id: UUID,
parent_run_id: UUID | None = None,
tags: list[str] | None = None,
metadata: dict[str, Any] | None = None,
**kwargs: Any,
) -> Any:
if metadata and (not tags or (TAG_NOSTREAM not in tags)):
ns = tuple(cast(str, metadata["langgraph_checkpoint_ns"]).split(NS_SEP))[
:-1
]
if not self.subgraphs and len(ns) > 0 and ns != self.parent_ns:
return
if tags:
if filtered_tags := [t for t in tags if not t.startswith("seq:step")]:
metadata["tags"] = filtered_tags
self.metadata[run_id] = (ns, metadata)
def on_llm_new_token(
self,
token: str,
*,
chunk: ChatGenerationChunk | None = None,
run_id: UUID,
parent_run_id: UUID | None = None,
tags: list[str] | None = None,
**kwargs: Any,
) -> Any:
if not isinstance(chunk, ChatGenerationChunk):
return
if meta := self.metadata.get(run_id):
self._emit(meta, chunk.message)
def on_llm_end(
self,
response: LLMResult,
*,
run_id: UUID,
parent_run_id: UUID | None = None,
**kwargs: Any,
) -> Any:
if meta := self.metadata.get(run_id):
if response.generations and response.generations[0]:
gen = response.generations[0][0]
if isinstance(gen, ChatGeneration):
self._emit(meta, gen.message, dedupe=True)
self.metadata.pop(run_id, None)
def on_llm_error(
self,
error: BaseException,
*,
run_id: UUID,
parent_run_id: UUID | None = None,
**kwargs: Any,
) -> Any:
self.metadata.pop(run_id, None)
def on_chain_start(
self,
serialized: dict[str, Any],
inputs: dict[str, Any],
*,
run_id: UUID,
parent_run_id: UUID | None = None,
tags: list[str] | None = None,
metadata: dict[str, Any] | None = None,
**kwargs: Any,
) -> Any:
if (
metadata
and kwargs.get("name") == metadata.get("langgraph_node")
and (not tags or TAG_HIDDEN not in tags)
):
ns = tuple(cast(str, metadata["langgraph_checkpoint_ns"]).split(NS_SEP))[
:-1
]
if not self.subgraphs and len(ns) > 0:
return
self.metadata[run_id] = (ns, metadata)
for value in _state_values(inputs):
if isinstance(value, BaseMessage):
if value.id is not None:
self.seen.add(value.id)
elif isinstance(value, Sequence) and not isinstance(value, str):
for item in value:
if isinstance(item, BaseMessage):
if item.id is not None:
self.seen.add(item.id)
def on_chain_end(
self,
response: Any,
*,
run_id: UUID,
parent_run_id: UUID | None = None,
**kwargs: Any,
) -> Any:
if meta := self.metadata.pop(run_id, None):
# Handle Command node updates
if isinstance(response, Command):
self._find_and_emit_messages(meta, response.update)
# Handle list of Command updates
elif isinstance(response, Sequence) and any(
isinstance(value, Command) for value in response
):
for value in response:
if isinstance(value, Command):
self._find_and_emit_messages(meta, value.update)
else:
self._find_and_emit_messages(meta, value)
# Handle basic updates / streaming
else:
self._find_and_emit_messages(meta, response)
def on_chain_error(
self,
error: BaseException,
*,
run_id: UUID,
parent_run_id: UUID | None = None,
**kwargs: Any,
) -> Any:
self.metadata.pop(run_id, None)
class StreamMessagesHandlerV2(StreamMessagesHandler, _V2StreamingCallbackHandler):
"""v2 variant of `StreamMessagesHandler`.
Declaring `_V2StreamingCallbackHandler` as a base flips
`BaseChatModel.invoke` to route through `_stream_chat_model_events`
(firing `on_stream_event`) instead of `_stream` (firing
`on_llm_new_token`). Inherits `on_stream_event` from the parent,
which forwards protocol events onto the messages stream channel.
Pregel attaches this class instead of the v1 handler only when
`StreamingHandler` opts in via the internal
`CONFIG_KEY_STREAM_MESSAGES_V2` config key; direct
`graph.stream(stream_mode="messages")` callers keep the v1
AIMessageChunk shape.
"""
def on_llm_new_token(
self,
token: str,
*,
chunk: ChatGenerationChunk | None = None,
run_id: UUID,
parent_run_id: UUID | None = None,
tags: list[str] | None = None,
**kwargs: Any,
) -> Any:
"""Intentional no-op — v1 chunks are not used on v2-flagged runs.
The v2 marker already steers `invoke` to the event generator, so
`on_llm_new_token` should not fire under normal routing. This
override stays a pass-through (no call to `super()`) to make
the intent explicit and to guard against any caller (e.g. a
node that calls `model.stream()` directly, which still fires
the v1 callback) leaking AIMessageChunks onto a v2-flagged
messages stream.
"""
# Intentionally empty: v2 handler does not forward v1 chunks.
def __init__(
self,
stream: Callable[[StreamChunk], None],
subgraphs: bool,
*,
parent_ns: tuple[str, ...] | None = None,
) -> None:
super().__init__(stream, subgraphs, parent_ns=parent_ns)
self._streamed_run_ids: set[UUID] = set()
def on_llm_end(
self,
response: LLMResult,
*,
run_id: UUID,
parent_run_id: UUID | None = None,
**kwargs: Any,
) -> Any:
if meta := self.metadata.get(run_id):
if response.generations and response.generations[0]:
gen = response.generations[0][0]
if isinstance(gen, ChatGeneration):
if run_id in self._streamed_run_ids:
if gen.message.id is None:
gen.message.id = str(uuid4())
self.seen.add(gen.message.id)
else:
self._emit(meta, gen.message, dedupe=True)
self._streamed_run_ids.discard(run_id)
self.metadata.pop(run_id, None)
def on_llm_error(
self,
error: BaseException,
*,
run_id: UUID,
parent_run_id: UUID | None = None,
**kwargs: Any,
) -> Any:
self._streamed_run_ids.discard(run_id)
super().on_llm_error(
error,
run_id=run_id,
parent_run_id=parent_run_id,
**kwargs,
)
def on_stream_event(
self,
event: dict[str, Any],
*,
run_id: UUID,
parent_run_id: UUID | None = None,
tags: list[str] | None = None,
**kwargs: Any,
) -> Any:
"""Forward a protocol event from `stream_events(version="v3")` as a messages stream part.
Fires once per `MessagesData` event (`message-start`, per-block
`content-block-*`, `message-finish`). The transformer layer
correlates events back to a single `ChatModelStream` via
`metadata["run_id"]` — attached here so the v1
`stream_mode="messages"` output (which emits
`(AIMessageChunk, metadata)` via `on_llm_new_token`) keeps its
original metadata shape.
Lives on the v2 handler rather than the v1 base: content-block
events are a v2-only concept, and forwarding them only when the
v2 handler is attached keeps the message channel's shape
predictable for v1 callers.
"""
if meta := self.metadata.get(run_id):
# Record message_id on message-start so on_chain_end's
# dedupe skips the finalized AIMessage the node returns
# (otherwise the messages projection double-counts: once
# from streaming, once from the chain output).
if event.get("event") == "message-start":
self._streamed_run_ids.add(run_id)
msg_id = event.get("message_id")
if msg_id:
self.seen.add(msg_id)
v2_meta = {**meta[1], "run_id": str(run_id)}
self.stream((meta[0], "messages", (event, v2_meta)))