同步已上线的智能分析执行协议、来源查询、图表报告与管理页面配套代码。 补齐数据库迁移、运行依赖和回归测试,保留原有指标计算口径。 统一午夜黑及其他主题的控件细节、中文和业务数字字体。 排除本地业务数据、凭据、预览、发布与验收产物。 验证:生产构建通过,320项前端回归通过;本轮仅同步Git,不再次部署。
144 lines
6.4 KiB
Python
144 lines
6.4 KiB
Python
"""Transport only thinking text explicitly returned by the configured provider."""
|
||
import copy
|
||
from datetime import datetime, timezone
|
||
import json
|
||
import re
|
||
import threading
|
||
import time
|
||
|
||
|
||
def public_thinking_text(text):
|
||
text = re.sub(r'quality-turn:[0-9a-f-]{36}', '', text, flags=re.I)
|
||
text = re.sub(r'(?i)Bearer\s+[A-Za-z0-9._~+/-]+', 'Bearer [REDACTED]', text)
|
||
text = re.sub(r'(?i)(api[_-]?key|password|token|secret|authorization)(\s*[=::]\s*)[^\s,,;;]+', r'\1\2[REDACTED]', text)
|
||
return re.sub(r'\bsk-[A-Za-z0-9_-]+', '[REDACTED]', text)
|
||
|
||
|
||
class ThinkingPublisher:
|
||
"""Publish full per-request snapshots; never derive thinking from an answer."""
|
||
def __init__(self, emit, interval=0.2, max_chars=1_000_000):
|
||
self.emit, self.interval, self.max_chars = emit, interval, max_chars
|
||
self.blocks, self.last, self.lock = {}, {}, threading.RLock()
|
||
self.closed = False
|
||
|
||
def receive(self, event):
|
||
with self.lock:
|
||
if self.closed:
|
||
return
|
||
key = event['id']
|
||
block = self.blocks.setdefault(key, {'id':key,'model':event['model'],
|
||
'phase':event['phase'],'status':'running','text':'','version':0,
|
||
'source':'api','streaming':event.get('streaming',True),'truncated':False,
|
||
'startedAt':datetime.now(timezone.utc).isoformat()})
|
||
if event['event'] == 'delta':
|
||
available = max(0, self.max_chars-len(block['text']))
|
||
delta = event.get('text','')
|
||
if delta and 'at' not in block:
|
||
block['at'] = datetime.now(timezone.utc).isoformat()
|
||
block['text'] += delta[:available]
|
||
block['truncated'] |= len(delta) > available
|
||
elif event['event'] == 'end':
|
||
block['finishedAt'] = datetime.now(timezone.utc).isoformat()
|
||
block['finishReason'] = event.get('finishReason','')
|
||
block['status'] = ('interrupted' if event.get('interrupted') else
|
||
'completed' if block['text'] else 'not_returned')
|
||
block['version'] += 1
|
||
if event['event'] != 'delta' or time.monotonic()-self.last.get(key,0) + 1e-9 >= self.interval:
|
||
self._publish(block)
|
||
|
||
def _publish(self, block):
|
||
public = copy.deepcopy(block)
|
||
public['text'] = public_thinking_text(public['text'])
|
||
self.last[block['id']] = time.monotonic()
|
||
self.emit(public)
|
||
|
||
def close(self):
|
||
with self.lock:
|
||
if self.closed:
|
||
return
|
||
self.closed = True
|
||
for block in self.blocks.values():
|
||
if block['status'] == 'running':
|
||
block['status'] = 'interrupted'
|
||
block['finishedAt'] = datetime.now(timezone.utc).isoformat()
|
||
block['version'] += 1
|
||
self._publish(block)
|
||
|
||
|
||
class ProviderStream:
|
||
"""Incremental SSE decoding plus a Chat Completions envelope for JSON callers."""
|
||
def __init__(self, on_thinking):
|
||
self.on_thinking = on_thinking
|
||
self.buffer = b''
|
||
self.choices, self.metadata, self.usage = {}, {}, None
|
||
self.tools = {}
|
||
self.done = False
|
||
|
||
def feed(self, chunk):
|
||
self.buffer += chunk
|
||
if len(self.buffer) > 2_000_000:
|
||
raise ValueError('Provider stream frame too large')
|
||
while b'\n' in self.buffer:
|
||
line, self.buffer = self.buffer.split(b'\n',1)
|
||
self._line(line.rstrip(b'\r'))
|
||
|
||
def finish(self):
|
||
if self.buffer:
|
||
self._line(self.buffer.rstrip(b'\r'))
|
||
self.buffer = b''
|
||
|
||
def _line(self, line):
|
||
if not line.startswith(b'data:'):
|
||
return
|
||
data = line[5:].strip()
|
||
if data == b'[DONE]':
|
||
self.done = True
|
||
return
|
||
if data:
|
||
self.document(json.loads(data))
|
||
|
||
def document(self, value):
|
||
if not isinstance(value,dict) or value.get('error'):
|
||
raise ValueError('Provider returned an invalid completion')
|
||
for field in ('id','model','created','system_fingerprint'):
|
||
if field in value:
|
||
self.metadata[field] = value[field]
|
||
if isinstance(value.get('usage'),dict):
|
||
self.usage = value['usage']
|
||
for choice in value.get('choices',[]):
|
||
index = choice.get('index',0)
|
||
target = self.choices.setdefault(index,{'index':index,'message':{'role':'assistant','content':''},'finish_reason':None})
|
||
delta = choice.get('delta') or choice.get('message') or {}
|
||
message = target['message']
|
||
for key in ('content','reasoning_content','reasoning'):
|
||
text = delta.get(key)
|
||
if isinstance(text,str):
|
||
message[key] = message.get(key,'') + text
|
||
# Aliases represent the same provider field; never double-display both.
|
||
thought = delta.get('reasoning_content') or delta.get('reasoning')
|
||
if isinstance(thought,str) and thought:
|
||
self.on_thinking(thought)
|
||
for position, call in enumerate(delta.get('tool_calls') or []):
|
||
tool_index = call.get('index',position)
|
||
tool = self.tools.setdefault(index,{}).setdefault(tool_index,{'type':'function','function':{'name':'','arguments':''}})
|
||
if isinstance(call.get('id'),str):tool['id']=call['id']
|
||
if isinstance(call.get('type'),str):tool['type']=call['type']
|
||
for field in ('name','arguments'):
|
||
part = (call.get('function') or {}).get(field)
|
||
if isinstance(part,str):tool['function'][field] += part
|
||
if choice.get('finish_reason') is not None:
|
||
target['finish_reason'] = choice['finish_reason']
|
||
|
||
def completion(self):
|
||
if not self.choices or any(not row['finish_reason'] for row in self.choices.values()):
|
||
raise ValueError('Provider stream ended before completion')
|
||
result = {**self.metadata,'object':'chat.completion','choices':list(self.choices.values())}
|
||
for choice in result['choices']:
|
||
tools = self.tools.get(choice['index'])
|
||
if tools:
|
||
choice['message']['tool_calls'] = [tools[key] for key in sorted(tools)]
|
||
if not choice['message']['content']:choice['message']['content']=None
|
||
if self.usage is not None:
|
||
result['usage'] = self.usage
|
||
return result
|