实战：如何使用图数据库提高向量搜索精确度？

URI: https://demo.neo4jlabs.com:7473/browser/
用户名: companies
密码: companies
数据库: companies

import os

os.environ["OPENAI_API_KEY"] = "sk-"
os.environ["NEO4J_URI"] = "neo4j+s://demo.neo4jlabs.com"
os.environ["NEO4J_USERNAME"] = "companies"
os.environ["NEO4J_PASSWORD"] = "companies"
os.environ["NEO4J_DATABASE"] = "companies"

embeddings = OpenAIEmbeddings()
graph = Neo4jGraph()
vector_index = Neo4jVector.from_existing_index(
 embeddings,
 index_name="news"
)

def get_organization_news(
 topic: Optional[str] = None,
 organization: Optional[str] = None,
 country: Optional[str] = None,
 sentiment: Optional[str] = None,
) -> str:
 # 如果没有预过滤条件，我们可以直接使用向量索引进行搜索
 if topic and not organization and not country and not sentiment:
     return vector_index.similarity_search(topic)
 # 使用并行运行时（如果可用）
 base_query = (
     "CYPHER runtime = parallel parallelRuntimeSupport=all "
     "MATCH (c:Chunk)<-[:HAS_CHUNK]-(a:Article) WHERE "
 )
 where_queries = []
 params = {"k": 5}  # 设置要检索的文本块数量

if organization:
    # 将组织名称映射到数据库中的候选项
    candidates = get_candidates(organization)
    if len(candidates) > 1:  # 如果候选选项太多，则需要用户进一步明确
        return f"请明确指出用户指的是以下哪个组织：{candidates}"
    # 添加一个过滤条件，筛选出提及特定组织的 articles
    where_queries.append(f"EXISTS {{(a)-[:MENTIONS]->(:Organization {{name: $organization}})}}")
    # 将组织名称作为参数传入
    params["organization"] = candidates[0]

if country:
    # 由于国家名称标准化，不需要额外的映射
    where_queries.append(f"EXISTS {{(a)-[:MENTIONS]->(:Organization)-[:IN_CITY]->()-[:IN_COUNTRY]->(:Country {{name: $country}})}}")
    params["country"] = country

if sentiment:
    if sentiment == "positive":
        where_queries.append("a.sentiment > $sentiment")
        params["sentiment"] = 0.5
    else:
        where_queries.append("a.sentiment < $sentiment")
        params["sentiment"] = -0.5

if topic:  # 执行向量比较
    vector_snippet = (
        "WITH c, a, vector.similarity.cosine(c.embedding,$embedding) AS score "
        "ORDER BY score DESC LIMIT toInteger($k)"
    )
    params["embedding"] = embeddings.embed_query(topic)
else:  # 只返回最新的数据
    vector_snippet = "WITH c, a ORDER BY a.date DESC LIMIT toInteger($k)"

return_snippet = "RETURN '#title ' + a.title + '\n#date ' + toString(a.date) + '\n#text ' + c.text AS output"

complete_query = (
    base_query + " AND ".join(where_queries) + vector_snippet + return_snippet
)

# 从数据库检索信息
data = graph.query(complete_query, params)
print(f"Cypher: {complete_query}\n")
# 在打印前安全地移除嵌入
params.pop('embedding', None)
print(f"参数: {params}")
return "###文章: ".join([el["output"] for el in data])

get_organization_news(
  organizatinotallow='neo4j',
  sentiment='positive',
  topic='远程工作'
)

# Cypher: CYPHER runtime = parallel parallelRuntimeSupport=all
# MATCH (c:Chunk)<-[:HAS_CHUNK]-(a:Article) WHERE 
# EXISTS {(a)-[:MENTIONS]->(:Organization {name: $organization})} AND 
# a.sentiment > $sentiment 
# WITH c, a, vector.similarity.cosine(c.embedding,$embedding) AS score 
# ORDER BY score DESC LIMIT toInteger($k) 
# RETURN '#title ' + a.title + '\n#date ' + toString(a.date) + '\n#text ' + c.text AS output

# 参数: {'k': 5, 'organization': 'Neo4j', 'sentiment': 0.5}

fewshot_examples = """{输入：Google员工的健康福利在新闻中有哪些？查询：健康福利}
{输入：关于Google的最新正面新闻是什么？查询：无}
{输入：有关VertexAI和Google的新闻有哪些？查询：VertexAI}
{输入：关于Google的新产品有哪些新闻？查询：新产品}
"""

class NewsInput(BaseModel):
    topic: Optional[str] = Field(
        descriptinotallow="除了组织、国家和情感倾向之外，如果您对其他特定信息或话题感兴趣，请告诉我们。以下是一些示例："
        + fewshot_examples
    )
    organization: Optional[str] = Field(
        descriptinotallow="您希望了解信息的组织名称"
    )
    country: Optional[str] = Field(
        descriptinotallow="您感兴趣的组织的所在国家。请使用正式的国家名称，例如‘美利坚合众国’或‘法国’。"
    )
    sentiment: Optional[str] = Field(
        descriptinotallow="您想要查询的文章情感倾向", enum=["正面", "负面"]
    )

class NewsTool(BaseTool):
    name = "新闻信息工具"
    description = (
        "当你需要在新闻中查找相关信息时，这个工具会非常有用。"
    )
    args_schema:Type[BaseModel] = NewsInput

    def _run(
        self,
        topic: Optional[str] = None,
        organization: Optional[str] = None,
        country: Optional[str] = None,
        sentiment: Optional[str] = None,
        run_manager: Optional[CallbackManagerForToolRun] = None,
    ) -> str:
        ""“使用这个工具来获取新闻信息。”""
        return get_organization_news(topic, organization, country, sentiment)

llm = ChatOpenAI(temperature=0, model="gpt-4-turbo", streaming=True)
tools = [NewsTool()]

llm_with_tools = llm.bind(functinotallow=[format_tool_to_openai_function(t) for t in tools])

prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            “你是一个乐于助人的助手，可以找到关于电影的信息并进行推荐。如果工具需要进一步的问题，请确保向用户询问以获得澄清。确保在后续问题中包含任何需要澄清的可用选项。只做用户明确请求的事情。”
        ),
        MessagesPlaceholder(variable_name="chat_history"),
        ("user", "{input}"),
        MessagesPlaceholder(variable_name="agent_scratchpad"),
    ]
)

agent = (
    {
        "input": lambda x: x["input"],
        "chat_history": lambda x: _format_chat_history(x["chat_history"])
        if x.get("chat_history")
        else [],
        "agent_scratchpad": lambda x: format_to_openai_function_messages(
            x["intermediate_steps"]
        ),
    }
    | prompt
    | llm_with_tools
    | OpenAIFunctionsAgentOutputParser()
)

agent_executor = AgentExecutor(agent=agent, tools=tools)

agent_executor.invoke(
  {"输入": "关于 neo4j 的一些正面新闻是什么？"}
)

# Cypher: CYPHER runtime = parallel parallelRuntimeSupport=all 
# MATCH (c:Chunk)<-[:HAS_CHUNK]-(a:Article) WHERE 
# EXISTS {(a)-[:MENTIONS]->(:Organization {name: $organization})} AND 
# a.sentiment > $sentiment WITH c, a 
# ORDER BY a.date DESC LIMIT toInteger($k) 
# RETURN '#标题 ' + a.title + '日期 ' + toString(a.date) + '文本 ' + c.text AS output
# 参数: {'k': 5, 'organization': 'Neo4j', 'sentiment': 0.5}

agent_executor.invoke(
   {"输入": "关于法国公司的员工幸福感，有哪些最新的负面新闻？"}
)

# Cypher: CYPHER runtime = parallel parallelRuntimeSupport=all 
# MATCH (c:Chunk)<-[:HAS_CHUNK]-(a:Article) WHERE 
# EXISTS {(a)-[:MENTIONS]->(:Organization)-[:IN_CITY]->()-[:IN_COUNTRY]->(:Country {name: $country})} AND 
# a.sentiment < $sentiment 
# WITH c, a, vector.similarity.cosine(c.embedding,$embedding) AS score 
# ORDER BY score DESC LIMIT toInteger($k) 
# RETURN '#标题 ' + a.title + '日期 ' + toString(a.date) + '文本 ' + c.text AS output
# 参数: {'k': 5, 'country': 'France', 'sentiment': -0.5, 'topic': '员工幸福感'}