You've already forked DataMate
feat(knowledge-graph): 实现知识图谱基础设施搭建
实现功能： - Neo4j Docker Compose 配置（社区版，端口 7474/7687，数据持久化） - Makefile 新增 Neo4j 命令（neo4j-up/down/logs/shell） - knowledge-graph-service Spring Boot 服务（完整的 DDD 分层架构） - kg_extraction Python 模块（基于 LangChain LLMGraphTransformer） 技术实现： - Neo4j 配置：环境变量化密码，统一默认值 datamate123 - Java 服务： - Domain: GraphEntity, GraphRelation 实体模型 - Repository: Spring Data Neo4j，支持 graphId 范围查询 - Service: 业务逻辑，graphId 双重校验，查询限流 - Controller: REST API，UUID 格式校验 - Exception: 实现 ErrorCode 接口，统一异常体系 - Python 模块： - KnowledgeGraphExtractor 类 - 支持异步/同步/批量抽取 - 支持 schema-guided 模式 - 兼容 OpenAI 及自部署模型 关键设计： - graphId 权限边界：所有实体操作都在正确的 graphId 范围内 - 查询限流：depth 和 limit 参数受配置约束 - 异常处理：统一使用 BusinessException + ErrorCode - 凭据管理：环境变量化，避免硬编码 - 双重防御：Controller 格式校验 + Service 业务校验 代码审查： - 经过 3 轮 Codex 审查和 2 轮 Claude 修复 - 所有 P0 和 P1 问题已解决 - 编译通过，无阻塞性问题 文件变更： - 新增：Neo4j 配置、knowledge-graph-service（11 个 Java 文件）、kg_extraction（3 个 Python 文件） - 修改：Makefile、pom.xml、application.yml、pyproject.toml
This commit is contained in:
17
runtime/datamate-python/app/module/kg_extraction/__init__.py
Normal file
17
runtime/datamate-python/app/module/kg_extraction/__init__.py
Normal file
@@ -0,0 +1,17 @@
|
||||
from app.module.kg_extraction.extractor import KnowledgeGraphExtractor
|
||||
from app.module.kg_extraction.models import (
|
||||
ExtractionRequest,
|
||||
ExtractionResult,
|
||||
Triple,
|
||||
GraphNode,
|
||||
GraphEdge,
|
||||
)
|
||||
|
||||
__all__ = [
|
||||
"KnowledgeGraphExtractor",
|
||||
"ExtractionRequest",
|
||||
"ExtractionResult",
|
||||
"Triple",
|
||||
"GraphNode",
|
||||
"GraphEdge",
|
||||
]
|
||||
183
runtime/datamate-python/app/module/kg_extraction/extractor.py
Normal file
183
runtime/datamate-python/app/module/kg_extraction/extractor.py
Normal file
@@ -0,0 +1,183 @@
|
||||
"""基于 LLM 的知识图谱三元组抽取器。
|
||||
|
||||
利用 LangChain 的 LLMGraphTransformer 从非结构化文本中抽取实体和关系,
|
||||
支持 schema-guided 抽取以提升准确率。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import logging
|
||||
from typing import Sequence
|
||||
|
||||
from langchain_core.documents import Document
|
||||
from langchain_openai import ChatOpenAI
|
||||
from langchain_experimental.graph_transformers import LLMGraphTransformer
|
||||
|
||||
from app.module.kg_extraction.models import (
|
||||
ExtractionRequest,
|
||||
ExtractionResult,
|
||||
ExtractionSchema,
|
||||
GraphEdge,
|
||||
GraphNode,
|
||||
Triple,
|
||||
)
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class KnowledgeGraphExtractor:
|
||||
"""基于 LLMGraphTransformer 的三元组抽取器。
|
||||
|
||||
Parameters
|
||||
----------
|
||||
model_name : str
|
||||
OpenAI 兼容模型名称。
|
||||
base_url : str | None
|
||||
自定义 API base URL（用于对接 vLLM/Ollama 等本地模型服务）。
|
||||
api_key : str
|
||||
API 密钥。
|
||||
temperature : float
|
||||
生成温度，抽取任务建议使用较低值。
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
model_name: str = "gpt-4o-mini",
|
||||
base_url: str | None = None,
|
||||
api_key: str = "EMPTY",
|
||||
temperature: float = 0.0,
|
||||
) -> None:
|
||||
self._llm = ChatOpenAI(
|
||||
model=model_name,
|
||||
base_url=base_url,
|
||||
api_key=api_key,
|
||||
temperature=temperature,
|
||||
)
|
||||
|
||||
def _build_transformer(
|
||||
self,
|
||||
schema: ExtractionSchema | None = None,
|
||||
) -> LLMGraphTransformer:
|
||||
"""根据可选的 schema 约束构造 LLMGraphTransformer。"""
|
||||
kwargs: dict = {"llm": self._llm}
|
||||
|
||||
if schema:
|
||||
if schema.entity_types:
|
||||
kwargs["allowed_nodes"] = [et.name for et in schema.entity_types]
|
||||
if schema.relation_types:
|
||||
kwargs["allowed_relationships"] = [rt.name for rt in schema.relation_types]
|
||||
|
||||
return LLMGraphTransformer(**kwargs)
|
||||
|
||||
async def extract(self, request: ExtractionRequest) -> ExtractionResult:
|
||||
"""从文本中抽取三元组。
|
||||
|
||||
Parameters
|
||||
----------
|
||||
request : ExtractionRequest
|
||||
包含文本、schema 约束等信息的抽取请求。
|
||||
|
||||
Returns
|
||||
-------
|
||||
ExtractionResult
|
||||
抽取得到的节点、边和三元组。
|
||||
"""
|
||||
transformer = self._build_transformer(request.schema)
|
||||
documents = [Document(page_content=request.text)]
|
||||
|
||||
try:
|
||||
graph_documents = await transformer.aconvert_to_graph_documents(documents)
|
||||
except Exception:
|
||||
logger.exception("LLM graph extraction failed for source_id=%s", request.source_id)
|
||||
return ExtractionResult(raw_text=request.text, source_id=request.source_id)
|
||||
|
||||
return self._convert_result(graph_documents, request)
|
||||
|
||||
def extract_sync(self, request: ExtractionRequest) -> ExtractionResult:
|
||||
"""同步版本的三元组抽取。"""
|
||||
transformer = self._build_transformer(request.schema)
|
||||
documents = [Document(page_content=request.text)]
|
||||
|
||||
try:
|
||||
graph_documents = transformer.convert_to_graph_documents(documents)
|
||||
except Exception:
|
||||
logger.exception("LLM graph extraction failed for source_id=%s", request.source_id)
|
||||
return ExtractionResult(raw_text=request.text, source_id=request.source_id)
|
||||
|
||||
return self._convert_result(graph_documents, request)
|
||||
|
||||
async def extract_batch(
|
||||
self,
|
||||
requests: Sequence[ExtractionRequest],
|
||||
) -> list[ExtractionResult]:
|
||||
"""批量抽取。
|
||||
|
||||
对多段文本逐一抽取并汇总结果。
|
||||
如需更高吞吐，可自行用 asyncio.gather 并发调用 extract。
|
||||
"""
|
||||
results: list[ExtractionResult] = []
|
||||
for req in requests:
|
||||
result = await self.extract(req)
|
||||
results.append(result)
|
||||
return results
|
||||
|
||||
@staticmethod
|
||||
def _convert_result(
|
||||
graph_documents: list,
|
||||
request: ExtractionRequest,
|
||||
) -> ExtractionResult:
|
||||
"""将 LangChain GraphDocument 转换为内部数据模型。"""
|
||||
nodes: list[GraphNode] = []
|
||||
edges: list[GraphEdge] = []
|
||||
triples: list[Triple] = []
|
||||
seen_nodes: set[str] = set()
|
||||
|
||||
for doc in graph_documents:
|
||||
# 收集节点
|
||||
for node in doc.nodes:
|
||||
node_key = f"{node.id}:{node.type}"
|
||||
if node_key not in seen_nodes:
|
||||
seen_nodes.add(node_key)
|
||||
nodes.append(
|
||||
GraphNode(
|
||||
name=node.id,
|
||||
type=node.type,
|
||||
properties=node.properties if hasattr(node, "properties") else {},
|
||||
)
|
||||
)
|
||||
|
||||
# 收集关系
|
||||
for rel in doc.relationships:
|
||||
source_node = GraphNode(
|
||||
name=rel.source.id,
|
||||
type=rel.source.type,
|
||||
)
|
||||
target_node = GraphNode(
|
||||
name=rel.target.id,
|
||||
type=rel.target.type,
|
||||
)
|
||||
|
||||
edges.append(
|
||||
GraphEdge(
|
||||
source=rel.source.id,
|
||||
target=rel.target.id,
|
||||
relation_type=rel.type,
|
||||
properties=rel.properties if hasattr(rel, "properties") else {},
|
||||
)
|
||||
)
|
||||
|
||||
triples.append(
|
||||
Triple(
|
||||
subject=source_node,
|
||||
predicate=rel.type,
|
||||
object=target_node,
|
||||
)
|
||||
)
|
||||
|
||||
return ExtractionResult(
|
||||
nodes=nodes,
|
||||
edges=edges,
|
||||
triples=triples,
|
||||
raw_text=request.text,
|
||||
source_id=request.source_id,
|
||||
)
|
||||
75
runtime/datamate-python/app/module/kg_extraction/models.py
Normal file
75
runtime/datamate-python/app/module/kg_extraction/models.py
Normal file
@@ -0,0 +1,75 @@
|
||||
"""知识图谱三元组抽取数据模型。"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from pydantic import BaseModel, Field
|
||||
|
||||
|
||||
class GraphNode(BaseModel):
|
||||
"""图谱节点（实体）。"""
|
||||
|
||||
name: str = Field(..., description="实体名称")
|
||||
type: str = Field(..., description="实体类型, 如 Person, Organization, Location")
|
||||
properties: dict[str, object] = Field(default_factory=dict, description="扩展属性")
|
||||
|
||||
|
||||
class GraphEdge(BaseModel):
|
||||
"""图谱边（关系）。"""
|
||||
|
||||
source: str = Field(..., description="源实体名称")
|
||||
target: str = Field(..., description="目标实体名称")
|
||||
relation_type: str = Field(..., description="关系类型, 如 works_at, located_in")
|
||||
properties: dict[str, object] = Field(default_factory=dict, description="关系属性")
|
||||
|
||||
|
||||
class Triple(BaseModel):
|
||||
"""知识三元组: (主体, 关系, 客体)。"""
|
||||
|
||||
subject: GraphNode
|
||||
predicate: str = Field(..., description="关系类型")
|
||||
object: GraphNode
|
||||
|
||||
|
||||
class EntityTypeConstraint(BaseModel):
|
||||
"""实体类型约束，用于 Schema-guided 抽取。"""
|
||||
|
||||
name: str = Field(..., description="类型名称")
|
||||
description: str = Field(default="", description="类型说明")
|
||||
|
||||
|
||||
class RelationTypeConstraint(BaseModel):
|
||||
"""关系类型约束。"""
|
||||
|
||||
name: str = Field(..., description="关系类型名称")
|
||||
source_types: list[str] = Field(default_factory=list, description="允许的源实体类型")
|
||||
target_types: list[str] = Field(default_factory=list, description="允许的目标实体类型")
|
||||
description: str = Field(default="", description="关系说明")
|
||||
|
||||
|
||||
class ExtractionSchema(BaseModel):
|
||||
"""抽取 schema 约束，约束 LLM 输出的实体和关系类型范围。"""
|
||||
|
||||
entity_types: list[EntityTypeConstraint] = Field(default_factory=list)
|
||||
relation_types: list[RelationTypeConstraint] = Field(default_factory=list)
|
||||
|
||||
|
||||
class ExtractionRequest(BaseModel):
|
||||
"""三元组抽取请求。"""
|
||||
|
||||
text: str = Field(..., description="待抽取的文本")
|
||||
graph_id: str = Field(..., description="目标图谱 ID")
|
||||
schema: ExtractionSchema | None = Field(
|
||||
default=None, description="可选的 schema 约束, 提供后做 schema-guided 抽取"
|
||||
)
|
||||
source_id: str | None = Field(default=None, description="来源 ID（数据集/知识库条目）")
|
||||
source_type: str = Field(default="KNOWLEDGE_BASE", description="来源类型")
|
||||
|
||||
|
||||
class ExtractionResult(BaseModel):
|
||||
"""三元组抽取结果。"""
|
||||
|
||||
nodes: list[GraphNode] = Field(default_factory=list)
|
||||
edges: list[GraphEdge] = Field(default_factory=list)
|
||||
triples: list[Triple] = Field(default_factory=list)
|
||||
raw_text: str = Field(default="", description="原始文本")
|
||||
source_id: str | None = None
|
||||
Reference in New Issue
Block a user