How to search your own files by meaning with Ollama, on your PC

Go to ollama.com/library/embeddinggemma-2. This model does not chat. It turns each file into a long list of numbers, and files with a similar meaning get similar numbers.

The page lists several sizes. We use 740m, which is 1.3 GB. The table says text and image, but it also accepted audio in our test. We tested only this size.

Open PowerShell and check Ollama. You should see a version number.

ollama --version

Then download the model. On our connection it took less than a minute.

ollama pull embeddinggemma-2:740m

When you read the word success, the model is on your computer.

Start with a test folder, not with your real documents.

mkdir my-files

Copy a few files into it. We used 13 sample files: six text notes, five photos, one screenshot and one voice memo. The photo names, such as IMG_2042.jpg, say nothing about what is inside.

Open Notepad and paste this script. Choose Save as, set the type to All files, and save it as find.py next to the my-files folder.

"""Search a folder by meaning. Everything stays on this PC."""
import base64
import json
import math
import sys
import time
import urllib.request
from pathlib import Path

URL = "http://localhost:11434/api/embed"
MODEL = "embeddinggemma-2:740m"
INDEX = Path("index.json")
TEXT = {".txt", ".md"}
IMAGES = {".jpg", ".jpeg", ".png"}
AUDIO = {".wav", ".mp3"}


def embed(item):
    body = json.dumps({"model": MODEL, "input": [item]}).encode()
    request = urllib.request.Request(URL, body, {"Content-Type": "application/json"})
    with urllib.request.urlopen(request, timeout=120) as reply:
        return json.load(reply)["embeddings"][0]


def similarity(a, b):
    top = sum(x * y for x, y in zip(a, b))
    return top / (math.sqrt(sum(x * x for x in a)) * math.sqrt(sum(y * y for y in b)))


def build(folder):
    start = time.time()
    files = []
    for path in sorted(Path(folder).rglob("*")):
        kind = path.suffix.lower()
        if kind in TEXT:
            content = path.read_text(encoding="utf-8", errors="ignore")[:8000]
            item = f"title: {path.stem} | text: {content}"
        elif kind in IMAGES:
            item = {"image": base64.b64encode(path.read_bytes()).decode()}
        elif kind in AUDIO:
            item = {"audio": base64.b64encode(path.read_bytes()).decode()}
        else:
            continue
        files.append({"path": str(path), "vector": embed(item)})
        print("Read", path.name)
    INDEX.write_text(json.dumps(files))
    print(f"Done: {len(files)} files in {time.time() - start:.1f} seconds")


def search(question):
    start = time.time()
    query = embed(f"task: search result | query: {question}")
    files = json.loads(INDEX.read_text())
    ranked = sorted(files, key=lambda f: similarity(query, f["vector"]), reverse=True)
    for f in ranked[:3]:
        print(f"{similarity(query, f['vector']):.2f}  {Path(f['path']).name}")
    print(f"Search took {time.time() - start:.2f} seconds")


if sys.argv[1] == "index":
    build(sys.argv[2])
else:
    search(" ".join(sys.argv[1:]))

Look at the line that starts with URL. localhost means this computer, so the script sends your files only to Ollama on your own PC. It does not use any online service. We read this in the code, but we did not run a test with the network unplugged.

In PowerShell, go to the folder that holds find.py, then run this command.

python find.py index my-files

Each file goes to the model, and the numbers are saved in a new file called index.json. Your original files do not change. The 13 files took 5.2 seconds here.

To remove the model, run this command.

ollama rm embeddinggemma-2:740m

Then delete the index. Your own files stay where they are.

Remove-Item index.json

To go further, our guide on how much GPU memory you need helps you choose hardware for local AI.

About the sample files. The train ticket is one we made, and the voice memo uses a generated voice. The five photos are CC0 images from Wikimedia Commons: lighthouse, dog, pizza, bicycle and mountains.