---
title: "Local Embedding Models · AssistantGeneral"
description: "Choose the model that turns documents into searchable meaning: multilingual EmbeddingGemma 2 by default, three others, background re-indexing, or a server."
url: "https://royalsoftworks.com/products/assistant-general/features/embedding-models/"
source: "https://royalsoftworks.com"
format: "markdown"
note: "Markdown rendering of the HTML page at `url`. Same content, same canonical URL."
---

Features · Your documents

# The model that reads your documents for search.

Choose the model that turns documents into searchable meaning: multilingual EmbeddingGemma 2 by default, three others, background re-indexing, or a server.

[Editions](https://royalsoftworks.com/products/assistant-general/editions/): Every edition runs the default. Choosing another: Pro and up.

## Search models, feature by feature

Each comes with a small drawing of how it works. They are diagrams, not screenshots; numbers in them are examples.

- ### Multilingual by default

  One model reads text, code, pictures and sound, in more than 100 languages, so documents in any language are searched without a translation step.

  For your IT team: EmbeddingGemma 2 (Google DeepMind, Apache 2.0), 768 dimensions.

  Diagram: Text and code and Pictures and sound and 100+ languages come together as One search space.
- ### Three alternatives

  An English-only model and two multilingual ones, a large and a small.

  For your IT team: bge-base-en-v1.5, multilingual-e5-large, multilingual-e5-small.

  Diagram: EmbeddingGemma 2, bge-base (English), e5-large, e5-small.
- ### Precision you choose

  Smaller and faster, or larger and more exact, per model.

  Scale from Smaller, faster to Larger, exact, set at “Your choice”.
- ### Switch without losing anything

  Changing the model re-indexes every document in the background from text already stored; the old model keeps answering until the new one is ready.

  Timeline (Old model answers until the new one is ready): Switch, Re-index, …, Ready.
- ### Your own embedding server

  Send the work to a dedicated server by address, model name and key, after a test that shows its vector size and speed.

  Diagram: Point it at an embedding server and test it: vector size and speed (768 wide and 35 ms are examples).
- ### Translate first, if you prefer

  With an English-only model, other languages can be translated before indexing.

  Diagram: Serbian text → Translate → English model.
- ### Never the same text twice

  A cache remembers what has been indexed, so unchanged text is never processed again.

  Diagram: Text, then “Seen before?”: yes, From cache; no, Index it.
- ### The company's model

  On a managed desk, document text can be indexed by the company server instead of the desk.

  Diagram: Desk → Company server → Vectors back.

## Goes well with

- [Document search](https://royalsoftworks.com/products/assistant-general/features/document-search/)
- [Recordings and video](https://royalsoftworks.com/products/assistant-general/features/audio-and-video/)
- [Extra databases](https://royalsoftworks.com/products/assistant-general/features/extra-databases/)

[Previous Database priority](https://royalsoftworks.com/products/assistant-general/features/database-priority/) [Next Files it reads](https://royalsoftworks.com/products/assistant-general/features/file-types/)

[All 31 feature classes](https://royalsoftworks.com/products/assistant-general/features/)

<!-- verse:royalsoftworks
     Isaija 55:6
     Tražite Gospoda, dok se može naći; prizivajte Ga, dokle je blizu.
     https://www.wordproject.org/bibles/verses/serbian/02_prayer.htm
-->