向量数据库

from langchain_community.vectorstores import FAISS
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_core.documents import Document

# 1. 示例文本和嵌入模型
texts = [
    "张三是法外狂徒",
    "FAISS是一个用于高效相似性搜索和密集向量聚类的库。",
    "LangChain是一个用于开发由语言模型驱动的应用程序的框架。"
]
docs = [Document(page_content=t) for t in texts]
embeddings = HuggingFaceEmbeddings(model_name="BAAI/bge-small-zh-v1.5")

# 2. 创建向量存储并保存到本地
vectorstore = FAISS.from_documents(docs, embeddings)

local_faiss_path = "./faiss_index_store"
vectorstore.save_local(local_faiss_path)

print(f"FAISS index has been saved to {local_faiss_path}")

# 3. 加载索引并执行查询
# 加载时需指定相同的嵌入模型，并允许反序列化
loaded_vectorstore = FAISS.load_local(
    local_faiss_path,
    embeddings,
    allow_dangerous_deserialization=True
)

# 相似性搜索
query = "FAISS是做什么的？"
results = loaded_vectorstore.similarity_search(query, k=1)

print(f"\n查询: '{query}'")
print("相似度最高的文档:")
for doc in results:
    print(f"- {doc.page_content}")

from llama_index.core import (
    Document,
    Settings,
    load_index_from_storage,
    StorageContext,
)
from llama_index.embeddings.huggingface import HuggingFaceEmbedding

# 1. 配置全局嵌入模型
Settings.embed_model = HuggingFaceEmbedding("BAAI/bge-small-zh-v1.5")

# 2. 创建示例文档
texts = [
    "张三是法外狂徒",
    "LlamaIndex是一个用于构建和查询私有或领域特定数据的框架。",
    "它提供了数据连接、索引和查询接口等工具。",
]
docs = [Document(text=t) for t in texts]

# 3. 创建索引并持久化到本地
index = VectorStoreIndex.from_documents(docs)
persist_path = "./llamaindex_index_store"
index.storage_context.persist(persist_dir=persist_path)
print(f"LlamaIndex 索引已保存至: {persist_path}")

# 4. 加载索引
loaded_index = StorageContext.from_defaults(persist_dir=persist_path)
vector_idnex = load_index_from_storage(loaded_index)

# 5. 构建查询
# 设置LlamaIndex的配置,使用本地Ollama大模型
from llama_index.llms.ollama import Ollama

Settings.llm = Ollama(
    model="gemma3-4b:Q4_K_M", base_url="http://127.0.0.1:11434", request_timeout=150
)
query = "LlamaIndex是做什么的？"
query_engine = vector_idnex.as_query_engine()
response = query_engine.query(query)
print(f"\n查询: '{query}'")
print(f"回答: {response}")

检索结果:
  Top 1: ID=460393204701700350, 距离=0.9466, 路径='../../data/C3\dragon\query.png'
  Top 2: ID=460393204701700344, 距离=0.7443, 路径='../../data/C3\dragon\dragon02.png'
  Top 3: ID=460393204701700348, 距离=0.6851, 路径='../../data/C3\dragon\dragon06.png'
  Top 4: ID=460393204701700341, 距离=0.6772, 路径='../../data/C3\dragon\dragon001.png'
  Top 5: ID=460393204701700343, 距离=0.6760, 路径='../../data/C3\dragon\dragon003.png'
  Top 6: ID=460393204701700342, 距离=0.6459, 路径='../../data/C3\dragon\dragon002.png'
  Top 7: ID=460393204701700345, 距离=0.6049, 路径='../../data/C3\dragon\dragon03.png'
  Top 8: ID=460393204701700347, 距离=0.5360, 路径='../../data/C3\dragon\dragon05.png'

# 编码查询
def encode_query(self, image_path: str, text: str) -> list[float]:
	with torch.no_grad():
		query_emb = self.model.encode(image=image_path, text=text)
	return query_emb.tolist()[0]

# 编码图像
def encode_image(self, image_path: str) -> list[float]:
	with torch.no_grad():
		query_emb = self.model.encode(image=image_path)
	return query_emb.tolist()[0]
	
query_text = "一条龙"
# 查询编码使用的是图片+文本
query_vector = encoder.encode_query(image_path=query_image_path, text=query_text)
# 如果直接采用图片，不使用文本，则会显示距离为1
# query_vector = encoder.encode_image(image_path=query_image_path)

检索结果:
  Top 1: ID=460393204701700360, 距离=1.0000, 路径='../../data/C3\dragon\query.png'
  Top 2: ID=460393204701700354, 距离=0.7908, 路径='../../data/C3\dragon\dragon02.png'
  Top 3: ID=460393204701700351, 距离=0.7321, 路径='../../data/C3\dragon\dragon001.png'
  Top 4: ID=460393204701700358, 距离=0.6975, 路径='../../data/C3\dragon\dragon06.png'
  Top 5: ID=460393204701700353, 距离=0.6867, 路径='../../data/C3\dragon\dragon003.png'
  Top 6: ID=460393204701700355, 距离=0.6592, 路径='../../data/C3\dragon\dragon03.png'
  Top 7: ID=460393204701700352, 距离=0.6551, 路径='../../data/C3\dragon\dragon002.png'
  Top 8: ID=460393204701700357, 距离=0.5703, 路径='../../data/C3\dragon\dragon05.png'