Ruslan Magana Vsevolodovna

Ruslan Magana Vsevolodovna

Machine Learning Engineer & Data Scientist & Physicist

Build Embeddings and Semantic Search with Sentence Transformers

1 minute read

Keyword search matches letters; semantic search matches meaning. The trick is to turn each piece of text into a vector (embedding) so that similar meanings land close together — then a query finds its nearest neighbours. This tutorial builds a working semantic search engine with Sentence Transformers and FAISS in a few lines of Python.

How it works

Semantic search: documents encoded to vectors, indexed in FAISS, queried by nearest neighbours

Prerequisites

  • Python 3.10+ (no GPU needed for this example)

1. Install

python -m pip install -U sentence-transformers faiss-cpu

2. Embed a corpus

all-MiniLM-L6-v2 is small, fast, and produces 384-dimensional vectors — perfect for getting started.

from sentence_transformers import SentenceTransformer

docs = [
    "How do I reset my password?",
    "The invoice total looks incorrect.",
    "Where can I download my receipt?",
    "My account is locked after too many attempts.",
    "How do I change my billing address?",
]

model = SentenceTransformer("all-MiniLM-L6-v2")
embeddings = model.encode(docs, normalize_embeddings=True)   # shape (5, 384)
print(embeddings.shape)   # (5, 384)

Normalizing the vectors means a plain dot-product equals cosine similarity — convenient for search.

3. Build a FAISS index

import faiss, numpy as np

dim = embeddings.shape[1]
index = faiss.IndexFlatIP(dim)        # inner product == cosine on normalized vectors
index.add(np.asarray(embeddings, dtype="float32"))
print(index.ntotal)   # 5

def search(query, k=3):
    q = model.encode([query], normalize_embeddings=True).astype("float32")
    scores, ids = index.search(q, k)
    return [(docs[i], round(float(s), 3)) for s, i in zip(scores[0], ids[0])]

for hit in search("I forgot my login credentials"):
    print(hit)

Expected output:

('My account is locked after too many attempts.', 0.62)
('How do I reset my password?', 0.61)
('Where can I download my receipt?', 0.27)

Notice the top hits never share the words “forgot” or “login” — they match on meaning. That’s the whole point of embeddings, and it’s exactly what powers retrieval-augmented generation (RAG).

5. Scale up

  • Swap IndexFlatIP for IndexIVFFlat or IndexHNSWFlat when you have millions of vectors.
  • Persist with faiss.write_index(index, "corpus.faiss") and reload with faiss.read_index(...).
  • For a managed vector database, the same vectors drop straight into Milvus, Pinecone, or pgvector.

Common errors

  • faiss import fails — install faiss-cpu (or faiss-gpu on CUDA machines).
  • Poor results — make sure you encode the query with the same model, and normalize both sides.
  • Slow first run — the model downloads once, then is cached locally.

FAQ

Which embedding model should I use? all-MiniLM-L6-v2 for speed; all-mpnet-base-v2 for higher quality; multilingual-e5-base for many languages.

Is this the same as RAG? Retrieval is the first half of RAG — you retrieve relevant chunks by embedding similarity, then feed them to an LLM to answer.

Leave a comment

You may also enjoy