Add Search to Your Product for $0 With EmbeddingGemma 2

uv venv -p 3.12 .venv && source .venv/bin/activate
uv pip install "sentence-transformers[image]" numpy

# without uv:
# python3.12 -m venv .venv && source .venv/bin/activate
# pip install "sentence-transformers[image]" numpy

import torch
from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    "google/embeddinggemma-2",
    model_kwargs={"torch_dtype": torch.float32},
    config_kwargs={"vision_config": None, "audio_config": None},  # text only
)
print(model.device)                                   # mps:0 on a Mac, cuda:0 on NVIDIA, else cpu
print(sum(p.numel() for p in model.parameters()))     # 271002624

import json
import numpy as np
import torch
from sentence_transformers import SentenceTransformer

rows = [json.loads(line) for line in open("catalog.jsonl")]
docs = [f"title: {r['name']} | text: {r['tagline']}. {r['description']}" for r in rows]

model = SentenceTransformer(
    "google/embeddinggemma-2",
    model_kwargs={"torch_dtype": torch.float32},
    config_kwargs={"vision_config": None, "audio_config": None},
)
vectors = model.encode(docs, prompt="", batch_size=32, normalize_embeddings=True)
np.save("catalog-768.npy", vectors)   # shape (2218, 768)

q = model.encode("check if my startup idea has real demand before I build it",
                 prompt_name="query", normalize_embeddings=True)
scores = vectors @ q
for i in np.argsort(-scores)[:3]:
    print(round(float(scores[i]), 3), rows[i]["name"])

# 0.788 Demand Discovery AI
# 0.782 IdeaScope
# 0.78 Idea Launch

small = vectors[:, :256]
small = small / np.linalg.norm(small, axis=1, keepdims=True)

# or at encode time:
# model.encode(docs, prompt="", truncate_dim=256, normalize_embeddings=True)

import glob
model = SentenceTransformer(
    "google/embeddinggemma-2",
    model_kwargs={"torch_dtype": torch.float32},
    config_kwargs={"audio_config": None},   # text + images
)
files = sorted(glob.glob("screenshots/*"))
shots = np.stack([
    model.encode({"text": "<|image|>", "image": [f]}, normalize_embeddings=True)
    for f in files
])

q = model.encode("a table of sales leads with deal size", prompt_name="query", normalize_embeddings=True)
print(files[int(np.argmax(shots @ q))])

import json
import numpy as np
import torch
from http.server import BaseHTTPRequestHandler, HTTPServer
from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    "google/embeddinggemma-2",
    device="cpu",
    model_kwargs={"torch_dtype": torch.float32},
    config_kwargs={"vision_config": None, "audio_config": None},
)
rows = [json.loads(line) for line in open("catalog.jsonl")]
vectors = np.load("catalog-768.npy")

class Handler(BaseHTTPRequestHandler):
    def do_POST(self):
        body = json.loads(self.rfile.read(int(self.headers["Content-Length"])))
        q = model.encode(body["q"], prompt_name="query", normalize_embeddings=True)
        scores = vectors @ q
        top = np.argsort(-scores)[: body.get("k", 5)]
        out = [{"slug": rows[i]["slug"], "name": rows[i]["name"],
                "score": round(float(scores[i]), 3)} for i in top]
        data = json.dumps(out).encode()
        self.send_response(200)
        self.send_header("content-type", "application/json")
        self.end_headers()
        self.wfile.write(data)

HTTPServer(("127.0.0.1", 8092), Handler).serve_forever()

python search_server.py &

curl -s -X POST localhost:8092 -H 'content-type: application/json' \
  -d '{"q":"an offline ebook reader that reads aloud","k":3}'
# [{"slug": "strophe-reader", "name": "Strophe Reader", "score": 0.832}, ...]

export async function semanticSearch(q: string, k = 5) {
    const res = await fetch("http://127.0.0.1:8092", {
        method: "POST",
        headers: { "content-type": "application/json" },
        body: JSON.stringify({ q, k }),
    });
    return (await res.json()) as Array<{ slug: string; name: string; score: number }>;
}