LangChain中LLM无法可靠检索知识图谱节点/信息的求助

问题背景 #

我基于表格构建了一个简易知识图谱（与实际图谱存在相同问题），通过NetworkxEntityGraph及循环语句创建，代码如下：

依赖导入与LLM初始化 #

import pandas as pd
import numpy as np

import os
import time
from langchain.chains import GraphQAChain
from langchain_community.graphs.networkx_graph import NetworkxEntityGraph

import networkx as nx
from langchain.llms import OpenAI

from langchain.chains import RetrievalQA

# Manage the API keys
from dotenv import load_dotenv

load_dotenv(dotenv_path='./API_KEYs.env')

# Google
from langchain_google_genai import GoogleGenerativeAI
import google.generativeai as genai

GOOGLE_API_KEY = os.getenv('GOOGLE_API_KEY')
genai.configure(api_key=GOOGLE_API_KEY)
llm = GoogleGenerativeAI(model='gemini-pro', google_api_key=GOOGLE_API_KEY)

图谱构建代码 #

fake_data = {'entity_a': ['Joanie', 'Johny', 'I'],
            'entity_b': ['Chachi', 'Here', 'Jenn-ay'],
            'interaction': ['loves', 'Where_he_is', 'still_loves']}

fake_data = pd.DataFrame(fake_data)

G = NetworkxEntityGraph()

#### ---- Add nodes ---- ####
for id, row in fake_data.iterrows():
    G._graph.add_node(row['entity_a'])
    G._graph.add_node(row['entity_b'])

#### ---- Add edges ---- ####
for id, row in fake_data.iterrows():
    G._graph.add_edge(
        row['entity_a'],
        row['entity_b'],
        relation=row['interaction']
    )

问题现象 #

直接查询图谱可正常获取节点及关系：

neighbors = list(G._graph.neighbors("Joanie"))
related_edges = list(G._graph.edges("Joanie", data=True))
print("Neighbors:", neighbors)
print("Related Edges:", related_edges)

输出：

Neighbors: ['Chachi']
Related Edges: [('Joanie', 'Chachi', {'relation': 'loves'})]

但使用GraphQAChain调用LLM时，无法识别图谱内容：

chain = GraphQAChain.from_llm(
    llm=llm,
    graph=G,
    verbose=True
)

question = "Tell me about Joanie's relation to the other people in the graph."

chain.invoke(question)

输出：

> Entering new GraphQAChain chain...
Entities Extracted:
NONE
Full Context:


> Finished chain.
{'query': "Tell me about Joanie's relation to the other people in the graph.",
 'result': 'This context does not mention anything about Joanie or her relation to the other people in the graph, so I cannot answer this question from the provided context.'}

即使是简单问题“Is Joanie in this graph?”，LLM仍返回“I don't know”。尝试使用旧教程中的GraphRetriever，发现该类已在LangChain更新后移除；后续测试中，即使LLM提取到实体Joanie且上下文显示“Joanie loves Chachi”，仍返回“I don't know”。

解决方案 #

针对LangChain最新版本的调整，给出以下修复步骤：

1. 使用公开方法构建图谱，避免操作私有属性 #

NetworkxEntityGraph的_graph是私有属性，直接操作可能导致LangChain无法识别节点和关系元数据。改用公开的add_node和add_edge方法：

G = NetworkxEntityGraph()

for id, row in fake_data.iterrows():
    # 添加节点时指定实体类型，帮助LLM识别
    G.add_node(row['entity_a'], type="Person")
    G.add_node(row['entity_b'], type="Person")
    # 添加边并指定关系类型
    G.add_edge(row['entity_a'], row['entity_b'], relation=row['interaction'])

2. 自定义实体提取逻辑适配Gemini #

默认的实体提取器可能对Gemini的输出格式适配不佳，手动定义实体提取的Prompt和Chain：

from langchain.chains import LLMChain
from langchain.prompts import PromptTemplate

# 自定义实体提取Prompt，明确要求提取人物实体
entity_extraction_prompt = PromptTemplate(
    input_variables=["question"],
    template="Extract all person entities mentioned in the question. Return only the entities as a comma-separated list with no extra text.\nQuestion: {question}"
)
# 创建实体提取Chain
entity_extractor = LLMChain(llm=llm, prompt=entity_extraction_prompt)

# 初始化GraphQAChain时指定自定义实体提取器
chain = GraphQAChain.from_llm(
    llm=llm,
    graph=G,
    entity_extractor=entity_extractor,
    verbose=True
)

3. 重写GraphQAChain的上下文获取逻辑 #

默认的上下文获取可能未正确从Networkx图谱中提取关系数据，手动替换该方法：

from langchain.chains.graph_qa.base import GraphQAChain

def custom_get_context(self, entities):
    context_parts = []
    for entity in entities:
        # 获取实体的所有关联边
        edges = list(self.graph._graph.edges(entity, data=True))
        for edge in edges:
            # 格式化关系描述
            context_parts.append(f"{edge[0]} {edge[2]['relation']} {edge[1]}")
        # 同时添加实体存在的信息
        context_parts.append(f"{entity} exists in the graph.")
    return "\n".join(context_parts)

# 替换原有的上下文获取方法
GraphQAChain._get_context = custom_get_context

4. 验证图谱元数据正确性 #

添加代码确认LangChain能识别图谱中的实体和关系：

print("Registered entities:", G.get_entities())
print("Registered relationships:", G.get_relationships())

如果输出为空，需手动补充NetworkxEntityGraph的元数据注册逻辑，或确保add_node/add_edge方法正确更新元数据。

内容的提问来源于stack exchange，提问作者Topher