Embeddings and vector search for analysts: a practical guide with EmbeddingGemma 2, DuckDB and PostgreSQL

Keyword search finds "refund" but misses "I want my money back". Embeddings fix that: they turn text, and now images and audio, into lists of numbers where similar meanings sit close together. They power AI search, chatbots that answer from your documents, duplicate detection and clustering. In October Google released EmbeddingGemma 2, a free, open model small enough to run on a laptop or phone. This guide explains the idea and walks through a complete example with tools analysts already use.

Embeddings in one picture

Every document is converted to a vector once and stored. At search time, the question is converted too, and the database returns the stored vectors closest to it, usually measured with cosine similarity, a score from -1 to 1 where higher means more similar.

Step 1: create embeddings in Python

Install the libraries once with pip install -U sentence-transformers transformers. The first run downloads the model from Hugging Face.

from sentence_transformers import SentenceTransformer

# Text only: skip the vision and audio encoders to save memory (270M parameters)
model = SentenceTransformer(
    "google/embeddinggemma-2",
    config_kwargs={"vision_config": None, "audio_config": None},
)

tickets = [
    "I was charged twice for my order",
    "Can I get my money back for a broken item?",
    "How do I change my delivery address?",
    "The app crashes when I open my basket",
]

# Documents use the "Document" prompt; 256 dimensions; unit length for cosine
doc_vecs = model.encode(tickets, prompt_name="Document",
                        truncate_dim=256, normalize_embeddings=True)

query_vec = model.encode("refund request", prompt_name="SearchQuery",
                         truncate_dim=256, normalize_embeddings=True)

print(model.similarity(query_vec, doc_vecs))  # highest score = best match

Step 2: store and search

import duckdb

con = duckdb.connect("tickets.duckdb")
con.execute("CREATE OR REPLACE TABLE tickets (id INT, text VARCHAR, emb FLOAT[256])")
con.executemany("INSERT INTO tickets VALUES (?, ?, ?)",
                [(i, t, v.tolist()) for i, (t, v) in enumerate(zip(tickets, doc_vecs))])

top = con.execute("""
    SELECT text, array_cosine_similarity(emb, ?::FLOAT[256]) AS score
    FROM tickets
    ORDER BY score DESC
    LIMIT 3
""", [query_vec.tolist()]).fetchall()
print(top)

CREATE EXTENSION IF NOT EXISTS vector;

CREATE TABLE tickets (
  id   bigserial PRIMARY KEY,
  text text NOT NULL,
  emb  vector(256)
);

-- Fast approximate search for large tables
CREATE INDEX ON tickets USING hnsw (emb vector_cosine_ops);

-- $1 is the query vector from Python; <=> is cosine distance (lower = closer)
SELECT text, 1 - (emb <=> $1) AS similarity
FROM tickets
ORDER BY emb <=> $1
LIMIT 5;

How many dimensions do you need?

EmbeddingGemma 2 is trained so you can keep only the first part of each vector. Shorter vectors mean less storage and faster search, with a small loss in quality. Google's own benchmark results:

What analysts can do with this

Working with Python data tools? See Upgrading to pandas 3.0 and DuckDB 2.0. For which AI models to pair with your search, read The best AI models in October 2026.

Sources: the EmbeddingGemma 2 model card on Hugging Face (google/embeddinggemma-2) and Google DeepMind's announcement (6 October 2026); DuckDB array functions documentation; pgvector documentation. Benchmarks are Google's published results.

Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 28 October 2026. About me