How Vector Stores Enable Lightning‑Fast Semantic Search in LangChain

What Is a Vector Store?

A vector store (or vector database) is a specialized database designed to store high‑dimensional vectors and perform efficient similarity search. When you have millions of document chunks each represented by a 1536‑dimensional embedding, brute‑force distance calculation is infeasible, so vector stores use approximate nearest‑neighbor (ANN) algorithms such as HNSW or IVF to achieve orders‑of‑magnitude speedups with minimal loss of accuracy.

Core Workflow

Adding/Indexing : Pass Document objects and their embeddings to the store, often via a convenience method like from_documents that handles embedding and insertion in one step.

Querying : Convert a user question into a query embedding using an embedding model.

Searching : Call the store’s search interface (e.g., similarity_search) with the query vector; the store returns the top‑k most similar documents.

Vector Stores in LangChain

LangChain abstracts dozens of vector‑store back‑ends behind a unified VectorStore interface, allowing you to swap implementations without changing application code. Common stores include: FAISS: an open‑source, in‑memory library ideal for prototyping and small‑to‑medium workloads. Chroma: an open‑source vector DB built for AI applications, runnable as a local or in‑memory service. Pinecone and Weaviate: managed cloud services offering high availability and scalability.

Key Methods

similarity_search(query, k=4, **kwargs)

: Takes a query string, embeds it automatically, and returns the k most similar Document objects. similarity_search_by_vector(embedding, k=4, **kwargs): Same as above but accepts a pre‑computed embedding. max_marginal_relevance_search(...): Balances relevance with diversity to avoid overly homogeneous results.

Example: Chinese Vector Store and Search

This example demonstrates how to use a Chinese‑optimized embedding model ( shibing624/text2vec-base-chinese) with LangChain’s FAISS store, showing both plain similarity search and similarity search with scores.

from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from langchain.schema import Document
from langchain_community.vectorstores.faiss import DistanceStrategy
import faiss

# Initialize Chinese embedding model
embedder = HuggingFaceEmbeddings(
    model_name="shibing624/text2vec-base-chinese",
    model_kwargs={'device': 'cpu'},
    encode_kwargs={'normalize_embeddings': True}
)

# Prepare some Chinese documents
documents = [
    Document(page_content="北京是中国的首都，有着悠久的历史文化。"),
    Document(page_content="上海是中国最大的经济中心，是一座现代化的国际大都市。"),
    Document(page_content="杭州以西湖闻名，是著名的旅游城市。"),
    Document(page_content="深圳是中国改革开放的窗口，高科技产业发达。"),
    Document(page_content="吐鲁番以其独特的火焰山和葡萄沟著称。"),
]

# Create the vector store
db = FAISS.from_documents(documents, embedder, distance_strategy=DistanceStrategy.MAX_INNER_PRODUCT)

# Perform similarity search
query = "中国哪个城市的科技最发达？"
docs = db.similarity_search(query)
print("
搜索结果：")
for i, doc in enumerate(docs, 1):
    print(f"
{i}. {doc.page_content}")

# Search with scores
docs_and_scores = db.similarity_search_with_score(query)
print("
带分数的搜索结果：")
for i, (doc, score) in enumerate(docs_and_scores, 1):
    print(f"
{i}. 文档内容: {doc.page_content}")
    print(f"   相似度分数: {score}")

The store is configured with DistanceStrategy.MAX_INNER_PRODUCT, which uses the inner product of normalized vectors as the similarity metric. Because embeddings are normalized, the inner‑product range is [-1, 1], where larger values indicate higher similarity.

Score Interpretation

Higher scores mean the document is more similar to the query (the opposite of cosine distance).

In the example, the Shenzhen document receives the highest score (≈0.82), confirming it is most relevant to “which Chinese city is most technologically advanced?”.

The Hangzhou document receives the lowest score (≈0.33), indicating the least relevance.

Result ordering (by descending score) is 深圳 > 上海 > 北京 > 杭州, matching the intuitive expectation for the query.

Correct Value Ranges for Different Strategies

EUCLIDEAN ('l2') : distance range [0, 2]; smaller values are more similar; 0 = identical, √2 = orthogonal, 2 = opposite.

COSINE ('cosine') : distance range [0, 2]; same interpretation as Euclidean for normalized vectors.

MAX_INNER_PRODUCT ('inner') : similarity range [-1, 1]; larger values are more similar; 1 = identical, 0 = orthogonal, -1 = opposite.

References

How to: use a vector store to retrieve data – https://python.langchain.com/docs/how_to/vectorstore_retrieval