sync: 鍚屾灞€鍩熺綉鏈€鏂颁唬鐮?(source master @ b69568e4)

This commit is contained in:
gozeros committed 2026-10-21 14:51:31 +08:00
1 parent 4f7f07504a
commit 8bc8b0e082
1651 files changed
+278599 -11416

No files matched your search

@@ -8,13 +8,12 @@ answer, while making no network or model-provider request.
from __future__ import annotations
import asyncio
import codecs
import math
import re
import uuid
from collections.abc import AsyncIterator, Sequence
from typing import Any
import tiktoken
from langchain_core.callbacks import AsyncCallbackManagerForLLMRun, CallbackManagerForLLMRun
from langchain_core.language_models.chat_models import BaseChatModel
from langchain_core.messages import AIMessage, AIMessageChunk, BaseMessage
@@ -24,12 +23,38 @@ from langgraph.graph import END, START, StateGraph
from deerflow.agents.thread_state import ThreadState
_TOKEN_ENCODING = tiktoken.get_encoding("cl100k_base")
_LOCAL_TOKEN_PATTERN = re.compile(
r"\s*(?:[\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff]|\w{1,4}|[^\w\s])|\s+",
re.UNICODE,
)
_TARGET_CHUNKS_PER_SECOND = 10
_MAX_TOKENS_PER_CHUNK = 20
_MAX_TOKENS_PER_SECOND = 1_000
def _split_local_tokens(text: str) -> list[str]:
"""Split text into lossless, token-like units without loading model assets.
Fixed answers only need a stable unit for stream pacing. Using a model
tokenizer here is both model-specific and unsafe for offline deployments:
``tiktoken.get_encoding`` downloads its vocabulary when the host cache is
empty. This local approximation treats each CJK character, short word
fragment, or punctuation mark as one unit and attaches surrounding
whitespace so the emitted chunks reconstruct the configured answer
exactly. Limiting word fragments also prevents a long URL or identifier
from bypassing the configured pace as one oversized unit.
"""
if not text:
return []
tokens = _LOCAL_TOKEN_PATTERN.findall(text)
if "".join(tokens) != text:
# The expression is intentionally exhaustive, but keep a lossless
# standard-library-only fallback for unusual Unicode input.
return list(text)
return tokens
class FixedAnswerChatModel(BaseChatModel):
"""A local streaming model that only replays an administrator answer."""
@@ -82,7 +107,7 @@ class FixedAnswerChatModel(BaseChatModel):
if not answer:
yield ChatGenerationChunk(message=AIMessageChunk(content="", id=self.message_id))
return
token_ids = _TOKEN_ENCODING.encode(answer, disallowed_special=())
tokens = _split_local_tokens(answer)
tokens_per_second = max(
1,
min(_MAX_TOKENS_PER_SECOND, self.tokens_per_second),
@@ -94,14 +119,14 @@ class FixedAnswerChatModel(BaseChatModel):
math.ceil(tokens_per_second / _TARGET_CHUNKS_PER_SECOND),
),
)
decoder = codecs.getincrementaldecoder("utf-8")()
token_groups = [token_ids[:1], *(token_ids[offset : offset + chunk_size] for offset in range(1, len(token_ids), chunk_size))]
for index, chunk_token_ids in enumerate(token_groups):
token_groups = [
tokens[:1],
*(tokens[offset : offset + chunk_size] for offset in range(1, len(tokens), chunk_size)),
]
for index, chunk_tokens in enumerate(token_groups):
if index > 0:
await asyncio.sleep(len(chunk_token_ids) / tokens_per_second)
chunk_bytes = b"".join(_TOKEN_ENCODING.decode_single_token_bytes(token_id) for token_id in chunk_token_ids)
is_last = index == len(token_groups) - 1
content = decoder.decode(chunk_bytes, final=is_last)
await asyncio.sleep(len(chunk_tokens) / tokens_per_second)
content = "".join(chunk_tokens)
chunk_kwargs = (
{
"fixed_answer": True,