# Settings › Embeddings

> Embeddings are what topics, cannibalization, semantic search and the topic map are built from. This section sets whether a new crawl makes them, how much of each page the model reads, and where the model runs.

Describes GLLmolfar 0.9.22, checked 2026-11-01. Canonical page: https://www.getlooploop.com/gllmolfar/docs/settings-embeddings

- **Where:** Settings › Embeddings
- **Model:** One model on this computer, the same in every mode
- **Built from:** The text of each crawled page
- **Changes:** From the next crawl

## On the screen

![Settings › Embeddings in GLLmolfar: four mode cards under a term explanation, the test button, and the model's text, thread and device settings](https://www.getlooploop.com/images/gllmolfar/docs/settings-embeddings.webp)

Numbered parts of the screenshot:

1. **A term explained.** Hovering the dotted **Embeddings** heading: one sentence and **More in the GetLoopLoop glossary**. Under it, **Embeddings on for new crawls**.
2. **Modes.** **Standard**, **Extended**, **Long** and **Longest**: how much text per section and how many sections. Each card shows its result in our test and, once tested, its result on this computer.
3. **Test on this computer.** Measures every mode on the open crawl's longest pages. Nothing leaves the computer.
4. **Locked modes.** **Modes beyond standard are for signed-in accounts (free).** **Sign in with Google** stands beside it, and the cards carry a lock.
5. **The model's settings.** **Text per page** and **Long pages**, the same choice as the cards; then **CPU threads for the model** and **Where the model runs**.
6. **Test and tune CPU and GPU.** Measures the CPU and the GPU and sets threads and batch sizes for this computer.

## The four modes

Every mode runs the same model, so changing mode downloads nothing. A mode is a size of text per section and a number of sections per page, printed on the card as `tokens × sections`. **Our test** is a fixed set of pages: a sentence from the second half of each page had to find its own page, and the figure is how often that page was in the first five.

| Card | Size | What the card says | Our test: right page in the top 5 |
| --- | --- | --- | --- |
| Standard | 256 × 4 | The start of each of four sections. Fastest. | 71% |
| Extended | 512 × 4 | Twice the text of each section. The best match in our test. | 80% |
| Long | 1024 × 4 | Four times the text. Slower, no better in our test. | 79% |
| Longest | 2048 × 2 | Two sections read almost whole. Slowest. | 77% |

After **Test on this computer**, each card shows how many pages per second it embedded here, how many times slower than Standard it was, and how often it **found its page** among the open crawl's pages. Three marks can appear: **Most accurate here**, **Fastest here** and **Found fewer here**. The result is kept until the next test, with the site, the number of pages and the date.

## Fields

| Field | Options | What it does |
| --- | --- | --- |
| Embeddings on for new crawls | On or off. On by default. | Whether the toolbar's **Embeddings** box starts ticked. Without embeddings a crawl has no topics, cannibalization or semantic search. |
| Text per page | 256 tokens (faster), 512 tokens (more precise), 1024 tokens, 2048 tokens (slowest) | How much text of each section the model reads. |
| Long pages | Start of the page only (fastest), Up to 2 sections, Up to 4 sections (recommended), Up to 8 sections (slowest) | How many sections of a long page are read. |
| CPU threads for the model | From 1 to the number of cores. Half the cores by default. | How many processor threads the model uses. |
| Where the model runs | Balanced: the faster of CPU and GPU (recommended), CPU only, GPU (DirectML) | Balanced uses the CPU until a test shows the GPU is clearly faster. The GPU choice and the test exist on Windows only; elsewhere the model runs on the CPU. |
| Test and tune CPU and GPU | A button | Tries thread counts and batch sizes on the CPU, then the GPU, in about two minutes, and keeps the best for this computer. |

## States and messages

- **Not tested** — the app says: "Not tested on this computer yet" A card before the first test. Only our test's figure is shown.
- **Testing** — the app says: "Mode 2 of 4: 512 × 4…" The mode test is running, one mode after another.
- **Busy** — the app says: "Wait until the crawl and its embeddings finish, then test." A test cannot run beside a crawl.
- **No pages** — the app says: "Open a crawl with saved pages first: the test uses its longest pages." The mode test needs an open crawl.
- **Failed** — the app says: "The test stopped: …" The reason follows the colon. Settings are unchanged.
- **Device not tested** — the app says: "Not tested yet. Balanced mode uses the CPU until a test shows the GPU is clearly faster." Before **Test and tune CPU and GPU** has run.
- **Device testing** — the app says: "Testing thread counts and batch sizes on the CPU, then the GPU: about two minutes…" The device test is running.
- **Stays on CPU** — the app says: "Balanced mode stays on the CPU: the GPU is not 1.5× faster, and a built-in GPU takes its memory from the same RAM." The test's verdict, with the measured rates before it.
- **Uses GPU** — the app says: "Balanced mode uses the GPU: it is worth its memory here." The test's verdict, with the measured rates before it.
- **GPU failed** — the app says: "The GPU could not run the model here; CPU … texts/s. Balanced mode uses the CPU." The model runs on the CPU.

## Related pages

- [Topics](https://www.getlooploop.com/gllmolfar/docs/topics.md): The Topics tab: the site's pages grouped by meaning, with each topic's size, key terms, cohesion, internal linking, cannibal pairs and central page.
- [Cannibalization](https://www.getlooploop.com/gllmolfar/docs/cannibalization.md): The Cannibalization tab: groups of pages that say almost the same thing, the page to keep, the pages side by side and their text passage by passage.
- [Semantic search](https://www.getlooploop.com/gllmolfar/docs/semantic-search.md): The Semantic search tab: type a question and get the crawled pages closest to it in meaning, with the sections of each page that match best.
- [Topic map](https://www.getlooploop.com/gllmolfar/docs/topic-map.md): The Topic map tab: every embedded page as a dot, placed by meaning and coloured by topic, status, indexability or depth, with a card for each page.
- [Settings › Storage](https://www.getlooploop.com/gllmolfar/docs/storage.md): Settings › Storage: how much disk the saved crawls, the embedding model and downloaded updates take, per project, and the actions that free it.