The NebiusRetriever enables efficient similarity search using embeddings from Nebius AI Studio. It leverages high-quality embedding models to enable semantic search over documents. This retriever is optimized for scenarios where you need to perform similarity search over a collection of documents, but don’t need to persist the vectors to a vector database. It performs vector similarity search in-memory using matrix operations, making it efficient for medium-sized document collections.

​
Setup

​
Installation

The Nebius integration can be installed via pip: 
%pip install --upgrade langchain-nebius

​
Credentials

Nebius requires an API key that can be passed as an initialization parameter api_key or set as the environment variable NEBIUS_API_KEY. You can obtain an API key by creating an account on Nebius AI Studio. 
import getpass
import os

# Make sure you've set your API key as an environment variable
if "NEBIUS_API_KEY" not in os.environ:
    os.environ["NEBIUS_API_KEY"] = getpass.getpass("Enter your Nebius API key: ")

​
Instantiation

The NebiusRetriever requires a NebiusEmbeddings instance and a list of documents. Here’s how to initialize it: 
from langchain_core.documents import Document
from langchain_nebius import NebiusEmbeddings, NebiusRetriever

# Create sample documents
docs = [
    Document(
        page_content="Paris is the capital of France", metadata={"country": "France"}
    ),
    Document(
        page_content="Berlin is the capital of Germany", metadata={"country": "Germany"}
    ),
    Document(
        page_content="Rome is the capital of Italy", metadata={"country": "Italy"}
    ),
    Document(
        page_content="Madrid is the capital of Spain", metadata={"country": "Spain"}
    ),
    Document(
        page_content="London is the capital of the United Kingdom",
        metadata={"country": "UK"},
    ),
    Document(
        page_content="Moscow is the capital of Russia", metadata={"country": "Russia"}
    ),
    Document(
        page_content="Washington DC is the capital of the United States",
        metadata={"country": "USA"},
    ),
    Document(
        page_content="Tokyo is the capital of Japan", metadata={"country": "Japan"}
    ),
    Document(
        page_content="Beijing is the capital of China", metadata={"country": "China"}
    ),
    Document(
        page_content="Canberra is the capital of Australia",
        metadata={"country": "Australia"},
    ),
]

# Initialize embeddings
embeddings = NebiusEmbeddings()

# Create retriever
retriever = NebiusRetriever(
    embeddings=embeddings,
    docs=docs,
    k=3,  # Number of documents to return
)

​
Usage

​
Retrieve Relevant Documents

You can use the retriever to find documents related to a query: 
# Query for European capitals
query = "What are some capitals in Europe?"
results = retriever.invoke(query)

print(f"Query: {query}")
print(f"Top {len(results)} results:")
for i, doc in enumerate(results):
    print(f"{i + 1}. {doc.page_content} (Country: {doc.metadata['country']})")

Query: What are some capitals in Europe?
Top 3 results:
1. Paris is the capital of France (Country: France)
2. Berlin is the capital of Germany (Country: Germany)
3. Rome is the capital of Italy (Country: Italy)

​
Using get_relevant_documents

You can also use the get_relevant_documents method directly (though invoke is the preferred interface): 
# Query for Asian countries
query = "What are the capitals in Asia?"
results = retriever.get_relevant_documents(query)

print(f"Query: {query}")
print(f"Top {len(results)} results:")
for i, doc in enumerate(results):
    print(f"{i + 1}. {doc.page_content} (Country: {doc.metadata['country']})")

Query: What are the capitals in Asia?
Top 3 results:
1. Beijing is the capital of China (Country: China)
2. Tokyo is the capital of Japan (Country: Japan)
3. Canberra is the capital of Australia (Country: Australia)

​
Customizing Number of Results

You can adjust the number of results at query time by passing k as a parameter: 
# Query for a specific country, with custom k
query = "Where is France?"
results = retriever.invoke(query, k=1)  # Override default k

print(f"Query: {query}")
print(f"Top {len(results)} result:")
for i, doc in enumerate(results):
    print(f"{i + 1}. {doc.page_content} (Country: {doc.metadata['country']})")

Query: Where is France?
Top 1 result:
1. Paris is the capital of France (Country: France)

​
Async Support

NebiusRetriever supports async operations: 
import asyncio


async def retrieve_async():
    query = "What are some capital cities?"
    results = await retriever.ainvoke(query)

    print(f"Async query: {query}")
    print(f"Top {len(results)} results:")
    for i, doc in enumerate(results):
        print(f"{i + 1}. {doc.page_content} (Country: {doc.metadata['country']})")


await retrieve_async()

Async query: What are some capital cities?
Top 3 results:
1. Washington DC is the capital of the United States (Country: USA)
2. Canberra is the capital of Australia (Country: Australia)
3. Paris is the capital of France (Country: France)

​
Handling Empty Documents

# Create a retriever with empty documents
empty_retriever = NebiusRetriever(
    embeddings=embeddings,
    docs=[],
    k=2,  # Empty document list
)

# Test the retriever with empty docs
results = empty_retriever.invoke("What are the capitals of European countries?")
print(f"Number of results: {len(results)}")

Number of results: 0

​
Use within a chain

NebiusRetriever works seamlessly in LangChain RAG pipelines. Here’s an example of creating a simple RAG chain with the NebiusRetriever: 
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables import RunnablePassthrough
from langchain_nebius import ChatNebius

# Initialize LLM
llm = ChatNebius(model="meta-llama/Llama-3.3-70B-Instruct-fast")

# Create a prompt template
prompt = ChatPromptTemplate.from_template(
    """
Answer the question based only on the following context:

Context:
{context}

Question: {question}
"""
)


# Format documents function
def format_docs(docs):
    return "\n\n".join(doc.page_content for doc in docs)


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

# Run the chain
answer = rag_chain.invoke("What are three European capitals?")
print(answer)

Based on the context provided, three European capitals are:

1. Paris
2. Berlin
3. Rome

​
Creating a Search Tool

You can use the NebiusRetrievalTool to create a tool for agents: 
from langchain_nebius import NebiusRetrievalTool

# Create a retrieval tool
tool = NebiusRetrievalTool(
    retriever=retriever,
    name="capital_search",
    description="Search for information about capital cities around the world",
)

# Use the tool
result = tool.invoke({"query": "capitals in Europe", "k": 3})
print("Tool results:")
print(result)

Tool results:
Document 1:
Paris is the capital of France

Document 2:
Berlin is the capital of Germany

Document 3:
Rome is the capital of Italy

​
How It Works

The NebiusRetriever works by:
  1. During initialization:
    • It stores the provided documents
    • It uses the provided NebiusEmbeddings to compute embeddings for all documents
    • These embeddings are stored in memory for quick retrieval
  2. During retrieval (invoke or get_relevant_documents):
    • It embeds the query using the same embedding model
    • It computes similarity scores between the query embedding and all document embeddings
    • It returns the top-k documents sorted by similarity
This approach is efficient for medium-sized document collections, as it avoids the need for a separate vector database while still providing high-quality semantic search.

​
API reference

For more details about the Nebius AI Studio API, visit the Nebius AI Studio Documentation.