z2api / app /utils /sse_tool_handler.py
zhaoxiaozhao07's picture
feat(core): 增强 OpenAI API 处理逻辑和工具支持
b98a223
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
SSE Tool Handler - 处理工具调用的SSE流
基于 Z.AI 原生的 edit_index 和 edit_content 机制,更原生地处理工具调用
"""
import json
import re
import time
from typing import Dict, Any, Optional, Generator, List
from app.utils.helpers import debug_log
class SSEToolHandler:
def __init__(self, chat_id: str, model: str):
self.chat_id = chat_id
self.model = model
# 工具调用状态
self.has_tool_call = False
self.tool_call_usage = None # 工具调用的usage信息
self.content_index = 0
self.has_thinking = False
self.content_buffer = bytearray() # 使用字节数组提高性能
self.last_edit_index = 0 # 上次编辑的位置
# 工具调用解析状态
self.active_tools = {} # 活跃的工具调用 {tool_id: tool_info}
self.completed_tools = [] # 已完成的工具调用
self.tool_blocks_cache = {} # 缓存解析的工具块
def process_tool_call_phase(self, data: Dict[str, Any], is_stream: bool = True) -> Generator[str, None, None]:
"""
处理tool_call阶段
"""
if not self.has_tool_call:
self.has_tool_call = True
debug_log("🔧 进入工具调用阶段")
edit_content = data.get("edit_content", "")
edit_index = data.get("edit_index", 0)
if not edit_content:
return
# debug_log(f"📦 接收内容片段 [index={edit_index}]: {edit_content[:1000]}...")
# 更新内容缓冲区
self._apply_edit_to_buffer(edit_index, edit_content)
# 尝试解析和处理工具调用
yield from self._process_tool_calls_from_buffer(is_stream)
def _apply_edit_to_buffer(self, edit_index: int, edit_content: str):
"""
在指定位置替换/插入内容更新内容缓冲区
"""
edit_bytes = edit_content.encode('utf-8')
required_length = edit_index + len(edit_bytes)
# 扩展缓冲区到所需长度(如果需要)
if len(self.content_buffer) < edit_index:
# 如果edit_index超出当前缓冲区,用空字节填充
self.content_buffer.extend(b'\x00' * (edit_index - len(self.content_buffer)))
# 确保缓冲区足够长以容纳新内容
if len(self.content_buffer) < required_length:
self.content_buffer.extend(b'\x00' * (required_length - len(self.content_buffer)))
# 在指定位置替换内容(不是插入,而是覆盖)
end_index = edit_index + len(edit_bytes)
self.content_buffer[edit_index:end_index] = edit_bytes
# debug_log(f"📝 缓冲区更新 [index={edit_index}, 长度={len(self.content_buffer)}]")
def _process_tool_calls_from_buffer(self, is_stream: bool) -> Generator[str, None, None]:
"""
从内容缓冲区中解析和处理工具调用
"""
try:
# 解码内容并清理空字节
content_str = self.content_buffer.decode('utf-8', errors='ignore').replace('\x00', '')
yield from self._extract_and_process_tools(content_str, is_stream)
except Exception as e:
debug_log(f"📦 内容解析暂时失败,等待更多数据: {e}")
# 不抛出异常,继续等待更多数据
def _extract_and_process_tools(self, content_str: str, is_stream: bool) -> Generator[str, None, None]:
"""
从内容字符串中提取和处理工具调用
"""
# 查找所有 glm_block,包括不完整的
pattern = r'<glm_block\s*>(.*?)(?:</glm_block>|$)'
matches = re.findall(pattern, content_str, re.DOTALL)
for block_content in matches:
# 尝试解析每个块
yield from self._process_single_tool_block(block_content, is_stream)
def _process_single_tool_block(self, block_content: str, is_stream: bool) -> Generator[str, None, None]:
"""
处理单个工具块,支持增量解析
"""
try:
# 尝试修复和解析完整的JSON
fixed_content = self._fix_json_structure(block_content)
tool_data = json.loads(fixed_content)
metadata = tool_data.get("data", {}).get("metadata", {})
tool_id = metadata.get("id", "")
tool_name = metadata.get("name", "")
arguments_raw = metadata.get("arguments", "{}")
if not tool_id or not tool_name:
return
debug_log(f"🎯 解析完整工具块: {tool_name}(id={tool_id}), 参数: {arguments_raw}")
# 检查是否是新工具或更新的工具
yield from self._handle_tool_update(tool_id, tool_name, arguments_raw, is_stream)
except json.JSONDecodeError as e:
debug_log(f"📦 JSON解析失败: {e}, 尝试部分解析")
# JSON 不完整,尝试部分解析
yield from self._handle_partial_tool_block(block_content, is_stream)
except Exception as e:
debug_log(f"📦 工具块处理失败: {e}")
def _fix_json_structure(self, content: str) -> str:
"""
修复JSON结构中的常见问题
"""
if not content:
return content
# 计算括号平衡
open_braces = content.count('{')
close_braces = content.count('}')
# 如果闭括号多于开括号,移除多余的闭括号
if close_braces > open_braces:
excess = close_braces - open_braces
fixed_content = content
for _ in range(excess):
# 从右侧移除多余的闭括号
last_brace_pos = fixed_content.rfind('}')
if last_brace_pos != -1:
fixed_content = fixed_content[:last_brace_pos] + fixed_content[last_brace_pos + 1:]
return fixed_content
return content
def _handle_tool_update(self, tool_id: str, tool_name: str, arguments_raw: str, is_stream: bool) -> Generator[str, None, None]:
"""
处理工具的创建或更新 - 更可靠的参数完整性检查
"""
# 解析参数
try:
if isinstance(arguments_raw, str):
# 先处理转义和清理
cleaned_args = self._clean_arguments_string(arguments_raw)
arguments = json.loads(cleaned_args) if cleaned_args.strip() else {}
else:
arguments = arguments_raw
except json.JSONDecodeError:
debug_log(f"📦 参数解析失败,暂不处理: {arguments_raw}")
# 参数解析失败时,不创建或更新工具,等待更完整的数据
return
# 检查参数是否看起来完整(基本的完整性验证)
is_args_complete = self._is_arguments_complete(arguments, arguments_raw)
# 检查是否是新工具
if tool_id not in self.active_tools:
debug_log(f"🎯 发现新工具: {tool_name}(id={tool_id}), 参数完整性: {is_args_complete}")
self.active_tools[tool_id] = {
"id": tool_id,
"name": tool_name,
"arguments": arguments,
"arguments_raw": arguments_raw,
"status": "active",
"sent_start": False,
"last_sent_args": {}, # 跟踪上次发送的参数
"args_complete": is_args_complete,
"pending_send": True # 标记需要发送
}
# 只有在参数看起来完整时才发送工具开始信号
if is_stream and is_args_complete:
yield self._create_tool_start_chunk(tool_id, tool_name, arguments)
self.active_tools[tool_id]["sent_start"] = True
self.active_tools[tool_id]["last_sent_args"] = arguments.copy()
self.active_tools[tool_id]["pending_send"] = False
debug_log(f"📤 发送完整工具开始: {tool_name}(id={tool_id})")
else:
# 更新现有工具
current_tool = self.active_tools[tool_id]
# 检查是否有实质性改进
if self._is_significant_improvement(current_tool["arguments"], arguments,
current_tool["arguments_raw"], arguments_raw):
debug_log(f"🔄 工具参数有实质性改进: {tool_name}(id={tool_id})")
current_tool["arguments"] = arguments
current_tool["arguments_raw"] = arguments_raw
current_tool["args_complete"] = is_args_complete
# 如果之前没有发送过开始信号,且现在参数完整,发送开始信号
if is_stream and not current_tool["sent_start"] and is_args_complete:
yield self._create_tool_start_chunk(tool_id, tool_name, arguments)
current_tool["sent_start"] = True
current_tool["last_sent_args"] = arguments.copy()
current_tool["pending_send"] = False
debug_log(f"📤 发送延迟的工具开始: {tool_name}(id={tool_id})")
# 如果已经发送过开始信号,且参数有显著改进,发送参数更新
elif is_stream and current_tool["sent_start"] and is_args_complete:
if self._should_send_argument_update(current_tool["last_sent_args"], arguments):
yield self._create_tool_arguments_chunk(tool_id, arguments)
current_tool["last_sent_args"] = arguments.copy()
debug_log(f"📤 发送参数更新: {tool_name}(id={tool_id})")
def _is_arguments_complete(self, arguments: Dict[str, Any], arguments_raw: str) -> bool:
"""
检查参数是否看起来完整
"""
if not arguments:
return False
# 检查原始字符串是否看起来完整
if not arguments_raw or not arguments_raw.strip():
return False
# 检查是否有明显的截断迹象
raw_stripped = arguments_raw.strip()
# 如果原始字符串不以}结尾,可能是截断的
if not raw_stripped.endswith('}') and not raw_stripped.endswith('"'):
return False
# 检查是否有不完整的URL(常见的截断情况)
for key, value in arguments.items():
if isinstance(value, str):
# 检查URL是否看起来完整
if 'http' in value.lower():
# 如果URL太短或以不完整的域名结尾,可能是截断的
if len(value) < 10 or value.endswith('.go') or value.endswith('.goo'):
return False
# 检查其他可能的截断迹象
if len(value) > 0 and value[-1] in ['.', '/', ':', '=']:
# 以这些字符结尾可能表示截断
return False
return True
def _is_significant_improvement(self, old_args: Dict[str, Any], new_args: Dict[str, Any],
old_raw: str, new_raw: str) -> bool:
"""
检查新参数是否比旧参数有显著改进
"""
# 如果新参数为空,不是改进
if not new_args:
return False
if len(new_args) > len(old_args):
return True
# 检查值的改进
for key, new_value in new_args.items():
old_value = old_args.get(key, "")
if isinstance(new_value, str) and isinstance(old_value, str):
# 如果新值明显更长且更完整,是改进
if len(new_value) > len(old_value) + 5: # 至少长5个字符才算显著改进
return True
# 如果旧值看起来是截断的,新值更完整,是改进
if old_value.endswith(('.go', '.goo', '.com/', 'http')) and len(new_value) > len(old_value):
return True
# 检查原始字符串的改进
if len(new_raw) > len(old_raw) + 10: # 原始字符串显著增长
return True
return False
def _should_send_argument_update(self, last_sent: Dict[str, Any], new_args: Dict[str, Any]) -> bool:
"""
判断是否应该发送参数更新 - 更严格的标准
"""
# 如果参数完全相同,不发送
if last_sent == new_args:
return False
# 如果新参数为空但之前有参数,不发送(避免倒退)
if not new_args and last_sent:
return False
# 如果新参数有更多键,发送更新
if len(new_args) > len(last_sent):
return True
# 检查是否有值变得显著更完整
for key, new_value in new_args.items():
last_value = last_sent.get(key, "")
if isinstance(new_value, str) and isinstance(last_value, str):
# 只有在值显著增长时才发送更新(避免微小变化)
if len(new_value) > len(last_value) + 5:
return True
elif new_value != last_value and new_value: # 确保新值不为空
return True
return False
def _handle_partial_tool_block(self, block_content: str, is_stream: bool) -> Generator[str, None, None]:
"""
处理不完整的工具块,尝试提取可用信息
"""
try:
# 尝试提取工具ID和名称
id_match = re.search(r'"id":\s*"([^"]+)"', block_content)
name_match = re.search(r'"name":\s*"([^"]+)"', block_content)
if id_match and name_match:
tool_id = id_match.group(1)
tool_name = name_match.group(1)
# 尝试提取参数部分
args_match = re.search(r'"arguments":\s*"([^"]*)', block_content)
partial_args = args_match.group(1) if args_match else ""
debug_log(f"📦 部分工具块: {tool_name}(id={tool_id}), 部分参数: {partial_args[:50]}")
# 如果是新工具,先创建记录
if tool_id not in self.active_tools:
# 尝试解析部分参数为字典
partial_args_dict = self._parse_partial_arguments(partial_args)
self.active_tools[tool_id] = {
"id": tool_id,
"name": tool_name,
"arguments": partial_args_dict,
"status": "partial",
"sent_start": False,
"last_sent_args": {},
"args_complete": False,
"partial_args": partial_args
}
if is_stream:
yield self._create_tool_start_chunk(tool_id, tool_name, partial_args_dict)
self.active_tools[tool_id]["sent_start"] = True
self.active_tools[tool_id]["last_sent_args"] = partial_args_dict.copy()
else:
# 更新部分参数
self.active_tools[tool_id]["partial_args"] = partial_args
# 尝试更新解析的参数
new_partial_dict = self._parse_partial_arguments(partial_args)
if new_partial_dict != self.active_tools[tool_id]["arguments"]:
self.active_tools[tool_id]["arguments"] = new_partial_dict
except Exception as e:
debug_log(f"📦 部分块解析失败: {e}")
def _clean_arguments_string(self, arguments_raw: str) -> str:
"""
清理和标准化参数字符串,改进对不完整JSON的处理
"""
if not arguments_raw:
return "{}"
# 移除首尾空白
cleaned = arguments_raw.strip()
# 处理特殊值
if cleaned.lower() == "null":
return "{}"
# 处理转义的JSON字符串
if cleaned.startswith('{\\"') and cleaned.endswith('\\"}'):
# 这是一个转义的JSON字符串,需要反转义
cleaned = cleaned.replace('\\"', '"')
elif cleaned.startswith('"{\\"') and cleaned.endswith('\\"}'):
# 双重转义的情况
cleaned = cleaned[1:-1].replace('\\"', '"')
elif cleaned.startswith('"') and cleaned.endswith('"'):
# 简单的引号包围,去除外层引号
cleaned = cleaned[1:-1]
# 处理不完整的JSON字符串
cleaned = self._fix_incomplete_json(cleaned)
# 标准化空格(移除JSON中的多余空格,但保留字符串值中的空格)
try:
# 先尝试解析,然后重新序列化以标准化格式
parsed = json.loads(cleaned)
if parsed is None:
return "{}"
cleaned = json.dumps(parsed, ensure_ascii=False, separators=(',', ':'))
except json.JSONDecodeError:
# 如果解析失败,只做基本的空格清理
debug_log(f"📦 JSON标准化失败,保持原样: {cleaned[:50]}...")
return cleaned
def _fix_incomplete_json(self, json_str: str) -> str:
"""
修复不完整的JSON字符串
"""
if not json_str:
return "{}"
# 确保以{开头
if not json_str.startswith('{'):
json_str = '{' + json_str
# 处理不完整的字符串值
if json_str.count('"') % 2 != 0:
# 奇数个引号,可能有未闭合的字符串
json_str += '"'
# 确保以}结尾
if not json_str.endswith('}'):
json_str += '}'
return json_str
def _parse_partial_arguments(self, arguments_raw: str) -> Dict[str, Any]:
"""
解析不完整的参数字符串,尽可能提取有效信息
"""
if not arguments_raw or arguments_raw.strip() == "" or arguments_raw.strip().lower() == "null":
return {}
try:
# 先尝试清理字符串
cleaned = self._clean_arguments_string(arguments_raw)
result = json.loads(cleaned)
# 确保返回字典类型
return result if isinstance(result, dict) else {}
except json.JSONDecodeError:
pass
try:
# 尝试修复常见的JSON问题
fixed_args = arguments_raw.strip()
# 处理转义字符
if '\\' in fixed_args:
fixed_args = fixed_args.replace('\\"', '"')
# 如果不是以{开头,添加{
if not fixed_args.startswith('{'):
fixed_args = '{' + fixed_args
# 如果不是以}结尾,尝试添加}
if not fixed_args.endswith('}'):
# 计算未闭合的引号和括号
quote_count = fixed_args.count('"') - fixed_args.count('\\"')
if quote_count % 2 != 0:
fixed_args += '"'
fixed_args += '}'
return json.loads(fixed_args)
except json.JSONDecodeError:
# 尝试提取键值对
return self._extract_key_value_pairs(arguments_raw)
except Exception:
# 如果所有方法都失败,返回空字典
return {}
def _extract_key_value_pairs(self, text: str) -> Dict[str, Any]:
"""
从文本中提取键值对,作为最后的解析尝试
"""
result = {}
try:
# 使用正则表达式提取简单的键值对
import re
# 匹配 "key": "value" 或 "key": value 格式
pattern = r'"([^"]+)":\s*"([^"]*)"'
matches = re.findall(pattern, text)
for key, value in matches:
result[key] = value
# 匹配数字值
pattern = r'"([^"]+)":\s*(\d+)'
matches = re.findall(pattern, text)
for key, value in matches:
try:
result[key] = int(value)
except ValueError:
result[key] = value
# 匹配布尔值
pattern = r'"([^"]+)":\s*(true|false)'
matches = re.findall(pattern, text)
for key, value in matches:
result[key] = value.lower() == 'true'
except Exception:
pass
return result
def _complete_active_tools(self, is_stream: bool) -> Generator[str, None, None]:
"""
完成所有活跃的工具调用 - 处理待发送的工具
"""
tools_to_send = []
for tool_id, tool in self.active_tools.items():
# 如果工具还没有发送过且参数看起来完整,现在发送
if is_stream and tool.get("pending_send", False) and not tool.get("sent_start", False):
if tool.get("args_complete", False):
debug_log(f"📤 完成时发送待发送工具: {tool['name']}(id={tool_id})")
yield self._create_tool_start_chunk(tool_id, tool["name"], tool["arguments"])
tool["sent_start"] = True
tool["pending_send"] = False
tools_to_send.append(tool)
else:
debug_log(f"⚠️ 跳过不完整的工具: {tool['name']}(id={tool_id})")
tool["status"] = "completed"
self.completed_tools.append(tool)
debug_log(f"✅ 完成工具调用: {tool['name']}(id={tool_id})")
self.active_tools.clear()
if is_stream and (self.completed_tools or tools_to_send):
# 发送工具完成信号
yield self._create_tool_finish_chunk()
def process_other_phase(self, data: Dict[str, Any], is_stream: bool = True) -> Generator[str, None, None]:
"""
处理other阶段 - 检测工具调用结束和状态更新
"""
edit_content = data.get("edit_content", "")
edit_index = data.get("edit_index", 0)
usage = data.get("usage")
# 保存usage信息
if self.has_tool_call and usage:
self.tool_call_usage = usage
debug_log(f"💾 保存工具调用usage: {usage}")
# 如果有edit_content,继续更新内容缓冲区
if edit_content:
self._apply_edit_to_buffer(edit_index, edit_content)
# 继续处理可能的工具调用更新
yield from self._process_tool_calls_from_buffer(is_stream)
# 检测工具调用结束的多种标记
if self.has_tool_call and self._is_tool_call_finished(edit_content):
debug_log("🏁 检测到工具调用结束")
# 完成所有活跃的工具
yield from self._complete_active_tools(is_stream)
if is_stream:
debug_log("🏁 发送工具调用完成信号")
yield "data: [DONE]"
# 重置工具调用状态
self.has_tool_call = False
def _is_tool_call_finished(self, edit_content: str) -> bool:
"""
检测工具调用是否结束的多种标记
"""
if not edit_content:
return False
# 检测各种结束标记
end_markers = [
"null,", # 原有的结束标记
'"status": "completed"', # 状态完成标记
'"is_error": false', # 错误状态标记
]
for marker in end_markers:
if marker in edit_content:
debug_log(f"🔍 检测到结束标记: {marker}")
return True
# 检查是否所有工具都有完整的结构
if self.active_tools and '"status": "completed"' in self.content_buffer:
return True
return False
def _reset_all_state(self):
"""重置所有状态"""
self.has_tool_call = False
self.tool_call_usage = None
self.content_index = 0
self.content_buffer = bytearray()
self.last_edit_index = 0
self.active_tools.clear()
self.completed_tools.clear()
self.tool_blocks_cache.clear()
def _create_tool_start_chunk(self, tool_id: str, tool_name: str, initial_args: Dict[str, Any] = None) -> str:
"""创建工具调用开始的chunk,支持初始参数"""
# 使用提供的初始参数,如果没有则使用空字典
args_dict = initial_args or {}
args_str = json.dumps(args_dict, ensure_ascii=False)
chunk = {
"choices": [
{
"delta": {
"role": "assistant",
"content": None,
"tool_calls": [
{
"id": tool_id,
"type": "function",
"function": {"name": tool_name, "arguments": args_str},
}
],
},
"finish_reason": None,
"index": self.content_index,
"logprobs": None,
}
],
"created": int(time.time()),
"id": self.chat_id,
"model": self.model,
"object": "chat.completion.chunk",
"system_fingerprint": "fp_zai_001",
}
return f"data: {json.dumps(chunk, ensure_ascii=False)}\n\n"
def _create_tool_arguments_chunk(self, tool_id: str, arguments: Dict) -> str:
"""创建工具参数的chunk - 只包含参数更新,不包含函数名"""
chunk = {
"choices": [
{
"delta": {
"tool_calls": [
{
"id": tool_id,
"function": {"arguments": json.dumps(arguments, ensure_ascii=False)},
}
],
},
"finish_reason": None,
"index": self.content_index,
"logprobs": None,
}
],
"created": int(time.time()),
"id": self.chat_id,
"model": self.model,
"object": "chat.completion.chunk",
"system_fingerprint": "fp_zai_001",
}
return f"data: {json.dumps(chunk, ensure_ascii=False)}\n\n"
def _create_tool_finish_chunk(self) -> str:
"""创建工具调用完成的chunk"""
chunk = {
"choices": [
{
"delta": {"role": "assistant", "content": None, "tool_calls": []},
"finish_reason": "tool_calls",
"index": 0,
"logprobs": None,
}
],
"created": int(time.time()),
"id": self.chat_id,
"usage": self.tool_call_usage or None,
"model": self.model,
"object": "chat.completion.chunk",
"system_fingerprint": "fp_zai_001",
}
return f"data: {json.dumps(chunk, ensure_ascii=False)}\n\n"