ZeusDB is a high-performance vector database powered by Rust, offering advanced features like product quantization, persistent storage, and enterprise-grade logging.
This documentation shows how to use ZeusDB to bring enterprise-grade vector search capabilities to your LangChain applications.

​
Quick Start

​
Installation

pip install langchain-zeusdb

​
Getting Started

This example uses OpenAIEmbeddings, which requires an OpenAI API key - Get your OpenAI API key here If you prefer, you can also use this package with any other embedding provider (Hugging Face, Cohere, custom functions, etc.). 
pip install langchain-openai

import os
import getpass

os.environ['OPENAI_API_KEY'] = getpass.getpass('OpenAI API Key:')

​
Basic Usage

from langchain_zeusdb import ZeusDBVectorStore
from langchain_openai import OpenAIEmbeddings
from zeusdb import VectorDatabase

# Initialize embeddings
embeddings = OpenAIEmbeddings(model="text-embedding-3-small")

# Create ZeusDB index
vdb = VectorDatabase()
index = vdb.create(
    index_type="hnsw",
    dim=1536,
    space="cosine"
)

# Create vector store
vector_store = ZeusDBVectorStore(
    zeusdb_index=index,
    embedding=embeddings
)

# Add documents
from langchain_core.documents import Document

docs = [
    Document(page_content="ZeusDB is fast", metadata={"source": "docs"}),
    Document(page_content="LangChain is powerful", metadata={"source": "docs"}),
]

vector_store.add_documents(docs)

# Search
results = vector_store.similarity_search("fast database", k=2)
print(f"Found the following {len(results)} results:")
print(results)
Expected results: 
Found the following 2 results:
[Document(id='ea2b4f13-b0b7-4cef-bb91-0fc4f4c41295', metadata={'source': 'docs'}, page_content='ZeusDB is fast'), Document(id='33dc1e87-a18a-4827-a0df-6ee47eabc7b2', metadata={'source': 'docs'}, page_content='LangChain is powerful')]

​
Factory Methods

For convenience, you can create and populate a vector store in a single step:

​
Example 1: Create from texts (creates index and adds texts in one step)

vector_store_texts = ZeusDBVectorStore.from_texts(
    texts=["Hello world", "Goodbye world"],
    embedding=embeddings,
    metadatas=[{"source": "text1"}, {"source": "text2"}]
)

print("texts store count:", vector_store_texts.get_vector_count())         # -> 2
print("texts store peek:", vector_store_texts.zeusdb_index.list(2))        # [('id1', {...}), ('id2', {...})]

# Search the texts-based store
results = vector_store_texts.similarity_search("Hello", k=1)
print(f"Found in texts store: {results[0].page_content}")                  # -> "Hello world"
Expected results: 
texts store count: 2
texts store peek: [('e9c39b44-b610-4e00-91f3-bf652e9989ac', {'source': 'text1', 'text': 'Hello world'}), ('d33f210c-ed53-4006-a64a-a9eee397fec9', {'source': 'text2', 'text': 'Goodbye world'})]
Found in texts store: Hello world

​
Example 2: - Create from documents (creates index and adds documents in one step)

new_docs = [
    Document(page_content="Python is great", metadata={"source": "python"}),
    Document(page_content="JavaScript is flexible", metadata={"source": "js"}),
]

vector_store_docs = ZeusDBVectorStore.from_documents(
    documents=new_docs,
    embedding=embeddings
)

print("docs store count:", vector_store_docs.get_vector_count())           # -> 2
print("docs store peek:", vector_store_docs.zeusdb_index.list(2))          # [('id3', {...}), ('id4', {...})]

# Search the documents-based store
results = vector_store_docs.similarity_search("Python", k=1)
print(f"Found in docs store: {results[0].page_content}")                   # -> "Python is great"
Expected results: 
docs store count: 2
docs store peek: [('aab2d1c1-7e02-4817-8dd8-6fb03570bb6f', {'text': 'Python is great', 'source': 'python'}), ('9a8a82cb-0e70-456c-9db2-556e464de14e', {'text': 'JavaScript is flexible', 'source': 'js'})]
Found in docs store: Python is great

​
Advanced Features

ZeusDB’s enterprise-grade capabilities are fully integrated into the LangChain ecosystem, providing quantization, persistence, advanced search features and many other enterprise capabilities.

​
Memory-Efficient Setup with Quantization

For large datasets, use Product Quantization to reduce memory usage: 
# Create quantized index for memory efficiency
quantization_config = {
    'type': 'pq',
    'subvectors': 8,
    'bits': 8,
    'training_size': 10000
}

vdb = VectorDatabase()
index = vdb.create(
    index_type="hnsw",
    dim=1536,
    space="cosine",
    quantization_config=quantization_config
)

vector_store = ZeusDBVectorStore(
    zeusdb_index=index,
    embedding=embeddings
)
Please refer to our documentation for helpful configuration guidelines and recommendations for setting up quantization.

​
Persistence

ZeusDB persistence lets you save a fully populated index to disk and load it later with complete state restoration. This includes vectors, metadata, HNSW graph, and (if enabled) Product Quantization models. What gets saved:
  • Vectors & IDs
  • Metadata
  • HNSW graph structure
  • Quantization config, centroids, and training state (if PQ is enabled)

​
How to Save your vector store

# Save index
vector_store.save_index("my_index.zdb")

​
How to Load your vector store

# Load index
loaded_store = ZeusDBVectorStore.load_index(
    path="my_index.zdb",
    embedding=embeddings
)

# Verify after load
print("vector count:", loaded_store.get_vector_count())
print("index info:", loaded_store.info())
print("store peek:", loaded_store.zeusdb_index.list(2))

​
Notes

  • The path is a directory, not a single file. Ensure the target is writable.
  • Saved indexes are cross-platform and include format/version info for compatibility checks.
  • If you used PQ, both the compression model and state are preserved—no need to retrain after loading.
  • You can continue to use all vector store APIs (similarity_search, retrievers, etc.) on the loaded_store.
For further details (including file structure, and further comprehensive examples), see the documentation.

​
Advanced Search Options

Use these to control scoring, diversity, metadata filtering, and retriever integration for your searches.

​
Similarity search with scores

Returns (Document, raw_distance) pairs from ZeusDB — lower distance = more similar. If you prefer normalized relevance in [0, 1], use similarity_search_with_relevance_scores. 
# Similarity search with scores
results_with_scores = vector_store.similarity_search_with_score(
    query="machine learning",
    k=5
)

print(results_with_scores)
Expected results: 
[
  (Document(id='ac0eaf5b-9f02-4ce2-8957-c369a7262c61', metadata={'source': 'docs'}, page_content='LangChain is powerful'), 0.8218843340873718),
  (Document(id='faae3adf-7cf3-463c-b282-3790b096fa23', metadata={'source': 'docs'}, page_content='ZeusDB is fast'), 0.9140053391456604)
]

​
MMR search for diversity

MMR (Maximal Marginal Relevance) balances two forces: relevance to the query and diversity among selected results, reducing near-duplicate answers. Control the trade-off with lambda_mult (1.0 = all relevance, 0.0 = all diversity). 
# MMR search for diversity
mmr_results = vector_store.max_marginal_relevance_search(
    query="AI applications",
    k=5,
    fetch_k=20,
    lambda_mult=0.7  # Balance relevance vs diversity
)

print(mmr_results)

​
Search with metadata filtering

Filter results using document metadata you stored when adding docs 
# Search with metadata filtering
results = vector_store.similarity_search(
    query="database performance",
    k=3,
    filter={"source": "documentation"}
)
For supported metadata query types and operators, please refer to the documentation.

​
As a Retriever

Turning the vector store into a retriever gives you a standard LangChain interface that chains (e.g., RetrievalQA) can call to fetch context. Under the hood it uses your chosen search type (similarity or mmr) and search_kwargs. 
# Convert to retriever for use in chains
retriever = vector_store.as_retriever(
    search_type="mmr",
    search_kwargs={"k": 3, "lambda_mult": 0.8}
)

# Use with LangChain Expression Language (LCEL) - requires only langchain-core
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough
from langchain_openai import ChatOpenAI

def format_docs(docs):
    return "\n\n".join([d.page_content for d in docs])

template = """Answer the question based only on the following context:
{context}

Question: {question}
"""

prompt = ChatPromptTemplate.from_template(template)
llm = ChatOpenAI()

# Create a chain using LCEL
chain = (
    {"context": retriever | format_docs, "question": RunnablePassthrough()}
    | prompt
    | llm
    | StrOutputParser()
)

# Use the chain
answer = chain.invoke("What is ZeusDB?")
print(answer)
Expected results: 
ZeusDB is a fast database management system.

​
Async Support

ZeusDB supports asynchronous operations for non-blocking, concurrent vector operations. When to use async: web servers (FastAPI/Starlette), agents/pipelines doing parallel searches, or notebooks where you want non-blocking/concurrent retrieval. If you’re writing simple scripts, the sync methods are fine. Those are asynchronous operations - the async/await versions of the regular synchronous methods. Here’s what each one does:
  1. await vector_store.aadd_documents(documents) - Asynchronously adds documents to the vector store (async version of add_documents())
  2. await vector_store.asimilarity_search("query", k=5) - Asynchronously performs similarity search (async version of similarity_search())
  3. await vector_store.adelete(ids=["doc1", "doc2"]) - Asynchronously deletes documents by their IDs (async version of delete())
The async versions are useful when:
  • You’re building async applications (using asyncio, FastAPI, etc.)
  • You want non-blocking operations that can run concurrently
  • You’re handling multiple requests simultaneously
  • You want better performance in I/O-bound applications
For example, instead of blocking while adding documents: 
# Synchronous (blocking)
vector_store.add_documents(docs)  # Blocks until complete

# Asynchronous (non-blocking)
await vector_store.aadd_documents(docs)  # Can do other work while this runs
All operations support async/await: Script version (python my_script.py): 
import asyncio
from langchain_zeusdb import ZeusDBVectorStore
from langchain_openai import OpenAIEmbeddings
from langchain_core.documents import Document
from zeusdb import VectorDatabase

# Setup
embeddings = OpenAIEmbeddings(model="text-embedding-3-small")
vdb = VectorDatabase()
index = vdb.create(index_type="hnsw", dim=1536, space="cosine")
vector_store = ZeusDBVectorStore(zeusdb_index=index, embedding=embeddings)

docs = [
    Document(page_content="ZeusDB is fast", metadata={"source": "docs"}),
    Document(page_content="LangChain is powerful", metadata={"source": "docs"}),
]

async def main():
    # Add documents asynchronously
    ids = await vector_store.aadd_documents(docs)
    print("Added IDs:", ids)

    # Run multiple searches concurrently
    results_fast, results_powerful = await asyncio.gather(
        vector_store.asimilarity_search("fast", k=2),
        vector_store.asimilarity_search("powerful", k=2),
    )
    print("Fast results:", [d.page_content for d in results_fast])
    print("Powerful results:", [d.page_content for d in results_powerful])

    # Delete documents asynchronously
    deleted = await vector_store.adelete(ids=ids[:1])
    print("Deleted first doc:", deleted)

if __name__ == "__main__":
    asyncio.run(main())
Colab/Notebook/Jupyter version (top-level await): 
from langchain_zeusdb import ZeusDBVectorStore
from langchain_openai import OpenAIEmbeddings
from langchain_core.documents import Document
from zeusdb import VectorDatabase
import asyncio

# Setup
embeddings = OpenAIEmbeddings(model="text-embedding-3-small")
vdb = VectorDatabase()
index = vdb.create(index_type="hnsw", dim=1536, space="cosine")
vector_store = ZeusDBVectorStore(zeusdb_index=index, embedding=embeddings)

docs = [
    Document(page_content="ZeusDB is fast", metadata={"source": "docs"}),
    Document(page_content="LangChain is powerful", metadata={"source": "docs"}),
]

# Add documents asynchronously
ids = await vector_store.aadd_documents(docs)
print("Added IDs:", ids)

# Run multiple searches concurrently
results_fast, results_powerful = await asyncio.gather(
    vector_store.asimilarity_search("fast", k=2),
    vector_store.asimilarity_search("powerful", k=2),
)
print("Fast results:", [d.page_content for d in results_fast])
print("Powerful results:", [d.page_content for d in results_powerful])

# Delete documents asynchronously
deleted = await vector_store.adelete(ids=ids[:1])
print("Deleted first doc:", deleted)
Expected results: 
Added IDs: ['9c440918-715f-49ba-9b97-0d991d29e997', 'ad59c645-d3ba-4a4a-a016-49ed39514123']
Fast results: ['ZeusDB is fast', 'LangChain is powerful']
Powerful results: ['LangChain is powerful', 'ZeusDB is fast']
Deleted first doc: True

​
Monitoring and Observability

​
Performance Monitoring

# Get index statistics
stats = vector_store.get_zeusdb_stats()
print(f"Index size: {stats.get('total_vectors', '0')} vectors")
print(f"Dimension: {stats.get('dimension')} | Space: {stats.get('space')} | Index type: {stats.get('index_type')}")

# Benchmark search performance
performance = vector_store.benchmark_search_performance(
    query_count=100,
    max_threads=4
)
print(f"Search QPS: {performance.get('parallel_qps', 0):.0f}")

# Check quantization status
if vector_store.is_quantized():
    progress = vector_store.get_training_progress()
    print(f"Quantization training: {progress:.1f}% complete")
else:
    print("Index is not quantized")
Expected results: 
Index size: 2 vectors
Dimension: 1536 | Space: cosine | Index type: HNSW
Search QPS: 53807
Index is not quantized

​
Enterprise Logging

ZeusDB includes enterprise-grade structured logging that works automatically with smart environment detection: 
import logging

# ZeusDB automatically detects your environment and applies appropriate logging:
# - Development: Human-readable logs, WARNING level
# - Production: JSON structured logs, ERROR level
# - Testing: Minimal output, CRITICAL level
# - Jupyter: Clean readable logs, INFO level

# Operations are automatically logged with performance metrics
vector_store.add_documents(docs)
# Logs: {"operation":"vector_addition","total_inserted":2,"duration_ms":45}

# Control logging with environment variables if needed
# ZEUSDB_LOG_LEVEL=debug ZEUSDB_LOG_FORMAT=json python your_app.py
To learn more about the full features of ZeusDB’s enterprise logging capabilities please read the following documentation.

​
Configuration Options

​
Index Parameters

vdb = VectorDatabase()
index = vdb.create(
    index_type="hnsw",           # Index algorithm
    dim=1536,                    # Vector dimension
    space="cosine",              # Distance metric: cosine, l2, l1
    m=16,                        # HNSW connectivity
    ef_construction=200,         # Build-time search width
    expected_size=100000,        # Expected number of vectors
    quantization_config=None     # Optional quantization
)

​
Search Parameters

results = vector_store.similarity_search(
    query="search query",
    k=5,                         # Number of results
    ef_search=None,              # Runtime search width (auto if None)
    filter={"key": "value"}      # Metadata filter
)

​
Error Handling

The integration includes comprehensive error handling: 
try:
    results = vector_store.similarity_search("query")
    print(results)
except Exception as e:
    # Graceful degradation with logging
    print(f"Search failed: {e}")
    # Fallback logic here

​
Requirements

  • Python: 3.10 or higher
  • ZeusDB: 0.0.8 or higher
  • LangChain Core: 0.3.74 or higher

​
Installation from Source

git clone https://github.com/zeusdb/langchain-zeusdb.git
cd langchain-zeusdb/libs/zeusdb
pip install -e .

​
Use Cases

  • RAG Applications: High-performance retrieval for question answering
  • Semantic Search: Fast similarity search across large document collections
  • Recommendation Systems: Vector-based content and collaborative filtering
  • Embeddings Analytics: Analysis of high-dimensional embedding spaces
  • Real-time Applications: Low-latency vector search for production systems

​
Compatibility

​
LangChain Versions

  • LangChain Core: 0.3.74+

​
Distance Metrics

  • Cosine: Default, normalized similarity
  • Euclidean (L2): Geometric distance
  • Manhattan (L1): City-block distance

​
Embedding Models

Compatible with any embedding provider:
  • OpenAI (text-embedding-3-small, text-embedding-3-large)
  • Hugging Face Transformers
  • Cohere Embeddings
  • Custom embedding functions

​
Support

Making vector search fast, scalable, and developer-friendly.