fix: improve ai writing chat streaming and intent handling

This commit is contained in:
董宏凯 committed 2026-07-02 17:43:23 +08:00
1 parent 845c87e3ee
commit da856f2b58
5 files changed
+1633 -158

No files matched your search

@@ -17,7 +17,7 @@ from __future__ import annotations
import logging
import re
from typing import Any
from typing import Any, AsyncIterator
from .prompts.intent_router_prompts import (
INTENT_ROUTER_SYSTEM,
@@ -48,6 +48,73 @@ ACTION_CONFIDENCE = 0.5
_SECTION_DECISIONS = {"supplement", "loose", "delete"}
_THINK_RE = re.compile(r"<think>.*?</think>", re.S)
_WRITING_ACTION_VERB_RE = re.compile(
r"(新增|增加|添加|加入|加上|补充|补上|补个|补一个|生成|撰写|写一|写个|写成|扩写|缩写|精简|改写|润色|调整|修改|重写|删除|删掉|拆成|合并|挪到|移动|规划|重新规划|设计|安排)"
)
_WRITING_ACTION_TARGET_RE = re.compile(
r"(第[一二三四五六七八九十\d]+[章节段]|章节|章|小节|段落|段|结尾|开头|标题|大纲|内容|案例|表格|总结章|总结章节|结论|摘要|观点|部分)"
)
_INFO_QUESTION_RE = re.compile(r"(哪些|什么|为什么|是什么|有哪些|多少|是否)")
def _fallback_intent() -> dict:
return {
"intent": "clarify",
"action": None,
"payload": {},
"need_research": False,
"research_query": "",
"answer": "",
"confidence": 0.0,
"clarify": "我没有理解你的意思,能换个说法再试一次吗?也可以直接点击上方卡片里的按钮操作。",
}
def _looks_like_writing_change_request(text: str) -> bool:
text = (text or "").strip()
if _INFO_QUESTION_RE.search(text):
return False
return bool(_WRITING_ACTION_VERB_RE.search(text) and _WRITING_ACTION_TARGET_RE.search(text))
def _rule_based_intent(text: str, pause_point: str | None) -> dict | None:
"""Deterministic guard for writing-operation requests that look like Q&A."""
if not _looks_like_writing_change_request(text):
return None
if pause_point == "outline_confirm":
return {
"intent": "revise_outline",
"action": "re_outline",
"payload": {"outlineFeedback": text.strip()},
"need_research": False,
"research_query": "",
"answer": "",
"confidence": 1.0,
"clarify": "",
}
if pause_point == "draft_confirm":
return {
"intent": "revise_draft",
"action": "user_revise",
"payload": {"userRevisionNotes": text.strip()},
"need_research": False,
"research_query": "",
"answer": "",
"confidence": 1.0,
"clarify": "",
}
if pause_point == "review_confirm":
return {
"intent": "accept_review_with_user_notes",
"action": "accept_review",
"payload": {"userRevisionNotes": text.strip()},
"need_research": False,
"research_query": "",
"answer": "",
"confidence": 1.0,
"clarify": "",
}
return None
# ── 上下文快照 ────────────────────────────────────────────────────────────────
@@ -362,16 +429,10 @@ async def parse_user_intent(
"""
from .nodes._utils import _content_to_text, parse_json
fallback = {
"intent": "clarify",
"action": None,
"payload": {},
"need_research": False,
"research_query": "",
"answer": "",
"confidence": 0.0,
"clarify": "我没有理解你的意思,能换个说法再试一次吗?也可以直接点击上方卡片里的按钮操作。",
}
fallback = _fallback_intent()
rule = _rule_based_intent(text, pause_point)
if rule is not None:
return normalize_intent_result(rule, pause_point, values, interrupt_value)
messages = _build_messages(text, pause_point, values, interrupt_value)
@@ -407,3 +468,106 @@ async def parse_user_intent(
parsed = parse_json(raw_text, fallback)
return normalize_intent_result(parsed, pause_point, values, interrupt_value)
async def stream_user_intent(
*,
text: str,
pause_point: str | None,
values: dict,
interrupt_value: dict | None = None,
model_name: str | None = None,
model: Any | None = None,
) -> AsyncIterator[dict]:
"""流式解析用户自由文本。
事件契约:
- ``{"type": "chunk", "chunk": "...", "is_thinking": bool}``
- ``{"type": "result", "result": normalized_intent}``
解析阶段仍以结构化 JSON 为最终产物;流式输出用于前端实时展示模型的
思考/解析过程。``<think>...</think>`` 内容和 provider reasoning 字段都会作为
``is_thinking=True`` 分流,正文 JSON 只参与最终解析。
"""
from .nodes._utils import ThinkTagParser, _content_to_text, parse_json, reasoning_delta
fallback = _fallback_intent()
rule = _rule_based_intent(text, pause_point)
if rule is not None:
yield {
"type": "result",
"result": normalize_intent_result(rule, pause_point, values, interrupt_value),
}
return
messages = _build_messages(text, pause_point, values, interrupt_value)
def content_events(parser: ThinkTagParser, raw_parts: list[str], delta: str) -> list[dict]:
events = []
for is_thinking, part in parser.feed(delta):
if not is_thinking:
raw_parts.append(part)
events.append({"type": "chunk", "chunk": part, "is_thinking": is_thinking})
return events
async def run_model(m: Any) -> AsyncIterator[dict]:
parser = ThinkTagParser()
raw_parts: list[str] = []
if hasattr(m, "astream"):
async for chunk in m.astream(messages):
reasoning = reasoning_delta(chunk)
if reasoning:
yield {"type": "chunk", "chunk": reasoning, "is_thinking": True}
content = _content_to_text(getattr(chunk, "content", "") or "")
if content:
for event in content_events(parser, raw_parts, content):
yield event
else:
resp = await m.ainvoke(messages)
content = _content_to_text(getattr(resp, "content", "") or "")
if content:
for event in content_events(parser, raw_parts, content):
yield event
yield {"type": "_raw_text", "raw_text": "".join(raw_parts)}
raw_text = ""
if model is not None:
try:
async for event in run_model(model):
if event["type"] == "_raw_text":
raw_text = str(event.get("raw_text") or "")
else:
yield event
except Exception:
raw_text = ""
logger.exception("intent_router: injected stream model failed")
else:
from deerflow.models import create_chat_model
from .nodes._model_fallback import classify_exception, fallback_chain
chain = fallback_chain((model_name or "").strip() or None) or [None]
for candidate in chain:
try:
m = create_chat_model(name=candidate) if candidate else create_chat_model()
async for event in run_model(m):
if event["type"] == "_raw_text":
raw_text = str(event.get("raw_text") or "")
else:
yield event
break
except Exception as exc: # noqa: BLE001 — 换下一个模型继续
raw_text = ""
logger.warning(
"intent_router: stream model=%s failed (reason=%s): %s",
candidate, classify_exception(exc), exc,
)
continue
raw_text = _THINK_RE.sub("", raw_text or "").strip()
if not raw_text:
result = normalize_intent_result(fallback, pause_point, values, interrupt_value)
else:
parsed = parse_json(raw_text, fallback)
result = normalize_intent_result(parsed, pause_point, values, interrupt_value)
yield {"type": "result", "result": result}