78acc19a88
- 后端新增 /files/import-workspace 接口，将工作区文件上传至 MinIO 并返回预处理结果 - 后端新增 resolve_workspace_file_path 工具函数，校验文件存在性与类型 - 前端 FileUploadModal 新增工作区文件选择模式，支持目录浏览与多选文件 - 前端 knowledge_api 新增 importWorkspaceFiles API - 图谱抽取器并发上限从 20 提升至 1000 - 图谱配置面板浮动层级与样式微调（阴影、边框、告警文案精简） - 知识库类型 milvus 标签颜色从 red 改为 blue - 新增工作区导入后端单元测试
67 lines
2.5 KiB
Python
67 lines
2.5 KiB
Python
from __future__ import annotations
|
||
|
||
from typing import Any
|
||
|
||
import json_repair
|
||
|
||
from yuxi.models.chat import select_model
|
||
|
||
from .base import GraphExtractor
|
||
|
||
|
||
DEFAULT_TRIPLE_EXTRACTION_PROMPT = """请从下面文本中抽取实体和实体关系，返回严格 JSON，不要输出解释。
|
||
JSON 格式：
|
||
{
|
||
"relations": [
|
||
{
|
||
"source": {"text": "实体文本", "label": "实体类型", "attributes": [{"text": "属性值", "label": "属性名称"}]},
|
||
"target": {"text": "实体文本", "label": "实体类型", "attributes": [{"text": "属性值", "label": "属性名称"}]},
|
||
"text": "关系显示文本",
|
||
"label": "关系类型"
|
||
}
|
||
]
|
||
}
|
||
"""
|
||
|
||
SCHEMA_INSTRUCTION = """抽取 Schema 约束：
|
||
{schema}
|
||
"""
|
||
|
||
|
||
class LLMGraphExtractor(GraphExtractor):
|
||
extractor_type = "llm"
|
||
|
||
def validate_options(self) -> None:
|
||
if not self.options.get("model_spec"):
|
||
raise ValueError("LLM 抽取器需要 model_spec")
|
||
if self.options.get("prompt"):
|
||
raise ValueError("LLM 图谱抽取器不支持自定义完整 Prompt，请使用 schema 配置抽取约束")
|
||
concurrency_count = self.options.get("concurrency_count", 1)
|
||
try:
|
||
concurrency_count = int(concurrency_count)
|
||
except (TypeError, ValueError) as exc:
|
||
raise ValueError("LLM 抽取器 concurrency_count 必须是整数") from exc
|
||
if concurrency_count < 1 or concurrency_count > 1000:
|
||
raise ValueError("LLM 抽取器 concurrency_count 必须在 1 到 1000 之间")
|
||
if self.options.get("model_params") is not None and not isinstance(self.options["model_params"], dict):
|
||
raise ValueError("LLM 抽取器 model_params 必须是对象")
|
||
|
||
async def extract(self, text: str, *, chunk_metadata: dict[str, Any] | None = None) -> dict[str, Any]:
|
||
self.validate_options()
|
||
model = select_model(
|
||
model_spec=self.options["model_spec"],
|
||
timeout=60.0,
|
||
model_params=self.options.get("model_params") or {},
|
||
)
|
||
prompt = self._build_prompt(text)
|
||
response = await model.call(prompt, stream=False)
|
||
parsed = json_repair.loads(response.content if response else "")
|
||
return parsed
|
||
|
||
def _build_prompt(self, text: str) -> str:
|
||
extraction_prompt = DEFAULT_TRIPLE_EXTRACTION_PROMPT
|
||
schema = str(self.options.get("schema") or "").strip()
|
||
if schema:
|
||
extraction_prompt = f"{extraction_prompt}\n{SCHEMA_INSTRUCTION.format(schema=schema)}"
|
||
return f"{extraction_prompt}\n\n文本：\n{text}"
|