langchain-vector-stores (Python)

Overview

Vector stores are databases optimized for storing and searching high-dimensional vectors (embeddings). They enable semantic search by finding documents similar to a query based on vector similarity rather than keyword matching. Essential for RAG (Retrieval-Augmented Generation) systems.

Key Concepts

• Vector Database: Specialized database for storing and querying embeddings
• Similarity Search: Finding similar documents using vector distance metrics (cosine, euclidean)
• Metadata Filtering: Combining vector search with metadata filters
• Persistence: Some vector stores run in-memory, others persist to disk/cloud
• Scaling: Different stores have different scalability characteristics

Vector Store Selection Decision Table

Vector Store	Best For	Package	Persistence	Scalability	Key Features
FAISS	Local, high performance	langchain-community	Disk	Medium	Fast, CPU/GPU support, local
Chroma	Development, simplicity	langchain-chroma	Disk	Medium	Easy setup, local-first
Pinecone	Production, managed	langchain-pinecone	Cloud	High	Fully managed, auto-scaling
InMemory	Testing, prototyping	langchain-core	Memory only	Low	Simple, no setup, ephemeral
Weaviate	GraphQL, hybrid search	langchain-weaviate	Cloud/Self-hosted	High	GraphQL, hybrid search
Qdrant	High performance, filtering	langchain-qdrant	Cloud/Self-hosted	High	Fast, advanced filtering
PGVector	PostgreSQL users	langchain-postgres	PostgreSQL	High	PostgreSQL extension

When to Choose Each Store

Choose FAISS if:

• You need high performance local vector search
• You want to avoid external dependencies
• You have many vectors (millions) to search quickly

Choose Chroma if:

• You want simple local development
• You need easy persistence without complex setup
• You're building a prototype or small application

Choose Pinecone if:

• You're building production applications
• You want zero operational overhead
• You need auto-scaling and high availability

Choose InMemory Vector Store if:

• You're testing or prototyping
• Data persistence isn't needed
• You want the simplest possible setup

Code Examples

InMemory Vector Store (Simplest)

from langchain_core.vectorstores import InMemoryVectorStore
from langchain_openai import OpenAIEmbeddings

# In-memory vector store - great for testing
vectorstore = InMemoryVectorStore(OpenAIEmbeddings())

# Add documents
from langchain_core.documents import Document

docs = [
    Document(page_content="LangChain is a framework for LLM apps", metadata={"source": "docs"}),
    Document(page_content="Vector stores enable semantic search", metadata={"source": "docs"}),
    Document(page_content="Paris is the capital of France", metadata={"source": "wiki"}),
]
vectorstore.add_documents(docs)

# Similarity search
results = vectorstore.similarity_search("What is LangChain?", k=2)
for doc in results:
    print(doc.page_content)

# Search with score
results_with_score = vectorstore.similarity_search_with_score("LangChain", k=2)
for doc, score in results_with_score:
    print(f"Score: {score}, Content: {doc.page_content}")

FAISS Vector Store

from langchain_community.vectorstores import FAISS
from langchain_openai import OpenAIEmbeddings
from langchain_core.documents import Document
import faiss

embeddings = OpenAIEmbeddings()

# Create from documents
docs = [
    Document(page_content="Document 1 content", metadata={"id": 1}),
    Document(page_content="Document 2 content", metadata={"id": 2}),
]
vectorstore = FAISS.from_documents(docs, embeddings)

# Search
results = vectorstore.similarity_search("query", k=3)

# Save to disk
vectorstore.save_local("./faiss_index")

# Load from disk
loaded_store = FAISS.load_local(
    "./faiss_index",
    embeddings,
    allow_dangerous_deserialization=True  # Required for loading
)

# Alternative: Initialize with specific FAISS index
embedding_dim = len(embeddings.embed_query("test"))
index = faiss.IndexFlatL2(embedding_dim)
docstore = InMemoryDocstore()
vectorstore = FAISS(embeddings, index, docstore, {})

Chroma Vector Store

from langchain_chroma import Chroma
from langchain_openai import OpenAIEmbeddings

# Persistent Chroma (saves to disk)
vectorstore = Chroma.from_documents(
    documents=docs,
    embedding=OpenAIEmbeddings(),
    persist_directory="./chroma_db",
    collection_name="my-collection",
)

# Search with metadata filter
results = vectorstore.similarity_search(
    "query",
    k=3,
    filter={"category": "A"}
)

# Load existing collection
vectorstore = Chroma(
    persist_directory="./chroma_db",
    embedding_function=OpenAIEmbeddings(),
    collection_name="my-collection",
)

# Delete collection
vectorstore.delete_collection()

Pinecone Vector Store

from langchain_pinecone import PineconeVectorStore
from pinecone import Pinecone, ServerlessSpec
from langchain_openai import OpenAIEmbeddings
import os

# Initialize Pinecone
pc = Pinecone(api_key=os.getenv("PINECONE_API_KEY"))

# Create index if it doesn't exist
index_name = "my-index"
if index_name not in pc.list_indexes().names():
    pc.create_index(
        name=index_name,
        dimension=1536,  # OpenAI embedding dimensions
        metric="cosine",
        spec=ServerlessSpec(cloud="aws", region="us-east-1")
    )

index = pc.Index(index_name)

# Create vector store
vectorstore = PineconeVectorStore.from_documents(
    documents=docs,
    embedding=OpenAIEmbeddings(),
    index=index,
)

# Search with metadata filter
results = vectorstore.similarity_search(
    "query",
    k=3,
    filter={"topic": "tech"}
)

# Use as retriever
retriever = vectorstore.as_retriever(
    search_kwargs={"k": 5}
)
docs = retriever.get_relevant_documents("query")

Adding Documents Incrementally

from langchain_core.vectorstores import InMemoryVectorStore
from langchain_openai import OpenAIEmbeddings

vectorstore = InMemoryVectorStore(OpenAIEmbeddings())

# Add documents
vectorstore.add_documents([
    Document(page_content="Document 1", metadata={}),
])

# Add more later
vectorstore.add_documents([
    Document(page_content="Document 2", metadata={}),
    Document(page_content="Document 3", metadata={}),
])

# Or from texts
vectorstore.add_texts(
    texts=["Text 1", "Text 2"],
    metadatas=[{"source": "A"}, {"source": "B"}]
)

Using as a Retriever

from langchain_community.vectorstores import FAISS
from langchain_openai import OpenAIEmbeddings, ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain.chains.combine_documents import create_stuff_documents_chain
from langchain.chains import create_retrieval_chain

# Create vector store
vectorstore = FAISS.from_documents(documents, OpenAIEmbeddings())

# Convert to retriever
retriever = vectorstore.as_retriever(
    search_type="similarity",  # or "mmr"
    search_kwargs={"k": 4}
)

# Use in a chain
llm = ChatOpenAI()
prompt = ChatPromptTemplate.from_template("""
Answer based on context:
{context}

Question: {input}
""")

document_chain = create_stuff_documents_chain(llm, prompt)
retrieval_chain = create_retrieval_chain(retriever, document_chain)

result = retrieval_chain.invoke({"input": "What is LangChain?"})
print(result["answer"])

Maximum Marginal Relevance (MMR)

from langchain_core.vectorstores import InMemoryVectorStore
from langchain_openai import OpenAIEmbeddings

vectorstore = InMemoryVectorStore.from_texts(
    texts=["text1", "text2", "text3"],
    metadatas=[{}, {}, {}],
    embedding=OpenAIEmbeddings()
)

# MMR balances relevance and diversity
results = vectorstore.max_marginal_relevance_search(
    "query",
    k=3,
    fetch_k=10,  # Fetch 10 candidates, return 3 diverse results
    lambda_mult=0.5  # 0 = max diversity, 1 = max relevance
)

Boundaries

What Agents CAN Do

✅ Initialize vector stores

• Set up any supported vector store
• Configure with embeddings and connection details

✅ Add and query documents

• Add documents with metadata
• Perform similarity search
• Use metadata filters

✅ Persist and load

• Save vector stores to disk (FAISS, Chroma)
• Load existing vector stores
• Manage collections

✅ Use as retrievers

• Convert vector stores to retrievers
• Integrate with chains and agents
• Configure search parameters

What Agents CANNOT Do

❌ Mix embeddings from different models

• Cannot use different embedding models within same vector store
• Must use consistent embeddings

❌ Bypass provider limits

• Cannot exceed Pinecone index size limits
• Cannot bypass free tier restrictions

❌ Modify vector dimensions after creation

• Cannot change embedding dimensions once store is created
• Must recreate store with new embeddings

Gotchas

1. FAISS Deserialization Warning

# ❌ Will raise error
loaded_store = FAISS.load_local("./faiss_index", embeddings)

# ✅ Must explicitly allow deserialization
loaded_store = FAISS.load_local(
    "./faiss_index",
    embeddings,
    allow_dangerous_deserialization=True
)

Fix: Add allow_dangerous_deserialization=True when loading FAISS indices.

2. Import from Correct Packages

# ❌ OLD: Using langchain imports
from langchain.vectorstores import FAISS  # Deprecated!
from langchain.vectorstores import Chroma

# ✅ NEW: Use specific packages
from langchain_community.vectorstores import FAISS
from langchain_chroma import Chroma
from langchain_pinecone import PineconeVectorStore

Fix: Use provider-specific packages.

3. Pinecone Index Creation

# ❌ Index doesn't auto-create
from pinecone import Pinecone
pc = Pinecone(api_key=api_key)
index = pc.Index("nonexistent")  # Error!

# ✅ Check and create
if "my-index" not in pc.list_indexes().names():
    pc.create_index(
        name="my-index",
        dimension=1536,
        metric="cosine",
        spec=ServerlessSpec(cloud="aws", region="us-east-1")
    )

Fix: Create Pinecone index before using it.

4. Chroma Persistence

# ❌ Not persisting
vectorstore = Chroma.from_documents(
    docs,
    OpenAIEmbeddings()
)  # Ephemeral!

# ✅ Persist to disk
vectorstore = Chroma.from_documents(
    docs,
    OpenAIEmbeddings(),
    persist_directory="./chroma_db"
)

Fix: Specify persist_directory for persistence.

5. Dimension Mismatch

# ❌ Pinecone index has 1536 dimensions, using 512-dim embeddings
pc.create_index(name="idx", dimension=1536, metric="cosine")

vectorstore = PineconeVectorStore.from_documents(
    docs,
    OpenAIEmbeddings(model="text-embedding-3-small", dimensions=512),
    index=pc.Index("idx")
)  # Error: dimension mismatch!

# ✅ Match dimensions
embeddings = OpenAIEmbeddings()  # Default 1536
# Or create index with 512 dimensions

Fix: Ensure vector store and embedding dimensions match.

Links and Resources

Official Documentation

• LangChain Python Vector Stores
• FAISS
• Chroma
• Pinecone

Provider Documentation

• FAISS Library
• Chroma
• Pinecone
• Qdrant

Package Installation

# FAISS
pip install langchain-community faiss-cpu
# or faiss-gpu for GPU support

# Chroma
pip install langchain-chroma

# Pinecone
pip install langchain-pinecone pinecone-client

# Qdrant
pip install langchain-qdrant qdrant-client