How to beginner · 3 min read

How to search with FAISS

Q: How to search with FAISS

Use FAISS to build a vector index and perform similarity search by adding vectors to an index and querying it with a query vector. The IndexFlatL2 index supports exact nearest neighbor search using L2 distance, and you can retrieve the top-k closest vectors efficiently.

Quick answer

Use FAISS to build a vector index and perform similarity search by adding vectors to an index and querying it with a query vector. The IndexFlatL2 index supports exact nearest neighbor search using L2 distance, and you can retrieve the top-k closest vectors efficiently.

PREREQUISITES

Python 3.8+
pip install faiss-cpu
Basic knowledge of vector embeddings

Setup

Install the faiss-cpu package via pip to get FAISS functionality on CPU. Import necessary modules and prepare your environment.

bash

pip install faiss-cpu

Step by step

This example shows how to create a FAISS index, add vectors, and perform a similarity search to find the nearest neighbors.

python

import numpy as np
import faiss

# Create some sample data: 5 vectors of dimension 4
vectors = np.array([
    [1.0, 0.0, 0.0, 0.0],
    [0.0, 1.0, 0.0, 0.0],
    [0.0, 0.0, 1.0, 0.0],
    [0.0, 0.0, 0.0, 1.0],
    [1.0, 1.0, 0.0, 0.0]
], dtype='float32')

# Initialize a FAISS index for L2 distance (exact search)
index = faiss.IndexFlatL2(4)  # 4 is the vector dimension

# Add vectors to the index
index.add(vectors)

# Query vector
query = np.array([[1.0, 0.0, 0.0, 0.0]], dtype='float32')

# Search for the 3 nearest neighbors
k = 3
D, I = index.search(query, k)  # D = distances, I = indices

print("Indices of nearest neighbors:", I)
print("Distances to nearest neighbors:", D)

output

Indices of nearest neighbors: [[0 4 1]]
Distances to nearest neighbors: [[0. 1. 2.]]

Common variations

You can use other FAISS index types for approximate search like IndexIVFFlat for large datasets. GPU support is available with faiss-gpu. For cosine similarity, normalize vectors before indexing.

python

import numpy as np
import faiss

# Example: Normalize vectors for cosine similarity
vectors = np.array([[1, 2, 3], [4, 5, 6]], dtype='float32')
faiss.normalize_L2(vectors)

# Create index for inner product (cosine similarity after normalization)
index = faiss.IndexFlatIP(3)
index.add(vectors)

query = np.array([[1, 0, 0]], dtype='float32')
faiss.normalize_L2(query)

D, I = index.search(query, 1)
print("Nearest neighbor index:", I)
print("Similarity score:", D)

output

Nearest neighbor index: [[0]]
Similarity score: [[0.26726124]]

Troubleshooting

If you get a dimension mismatch error, verify your vectors and query have the same dimension.
For large datasets, use an approximate index like IndexIVFFlat and train it before adding vectors.
Ensure vectors are float32 numpy arrays; other types cause errors.

✅

Key Takeaways

Use faiss.IndexFlatL2 for exact nearest neighbor search with L2 distance.
Normalize vectors to use cosine similarity with IndexFlatIP.
For large datasets, use approximate indexes like IndexIVFFlat with training.
Always ensure vectors are numpy arrays of dtype float32.
FAISS supports GPU acceleration via faiss-gpu for faster search.

Verified 2026-04

Verify ↗