Embedded Models
The catalog of local Ollama embedding models available for download from Settings → Agentic, and how to pick the right one for your knowledge base.
Overview
Embedding models don't generate text — they convert text into a vector (a list of numbers) that captures its meaning, so that pieces of content with similar meaning end up with similar vectors. Semantic Search uses this to rank your knowledge base by relevance, rather than by exact keyword matches.
By default, embedding models run through Ollama — the same local server process on your machine, reachable over HTTP (a hosted alternative is described under Hosted (Cloud LLM) Embedding below). They are configured and downloaded separately from text-generation models, since a knowledge base is indexed with one embedding model and must keep using that same model for all future searches to stay consistent.
Where to configure it
Go to Settings → AI Models → Embedding Engine. From there you can check your Ollama connection status, select an already-installed embedding model, or download a new one directly from the built-in model list described below. See Ollama Integration for how TotalApp connects to Ollama, Agentic AI Settings for the full settings reference, and Embedding Model for how the selected model is used by Semantic Search.
Underlying Technology
Like the text-generation catalog, embedding models are served by Ollama, a standalone local server application built on llama.cpp:
- Models are downloaded in GGUF format and run through llama.cpp's embedding-generation path rather than its text-generation path.
- Runtime: llama.cpp, running as a local server process reachable over HTTP (
127.0.0.1:11434by default) — the same Ollama server used for text generation, just a different model and API endpoint.
See Local Language Models for the text-generation catalog served by the same Ollama server.
Two Great Starting Points
The Settings → Agentic download list is pre-loaded with the most widely used embedding models in the Ollama community. If you're not sure where to start, these two cover the two most common needs:
nomic-embed-text
A long-context specialist. Its 8192-token context window makes it the strongest default choice for English-language documents and long-form RAG (Retrieval-Augmented Generation) pipelines, where a short context window would otherwise truncate large sources.
bge-m3
A multilingual specialist — the "M3" in its name stands for Multi-Lingual, Multi-Functionality, Multi-Granularity. It supports 100+ languages and performs especially well on Turkish and other non-English text similarity tasks.
Together, nomic-embed-text and bge-m3 already cover the two most common scenarios — long English documents, and multilingual or Turkish content — which is why they're the first two entries in the download list.
Full Model Catalog
Beyond the two starting points above, the download list also offers models tuned for retrieval-heavy workloads, additional multilingual coverage, and ultra-lightweight use on low-powered machines. Approximate download sizes are shown next to each model in the Settings UI.
| Model | Size | Category | Best for |
|---|---|---|---|
nomic-embed-text:latest | ~0.3 GB | English & Retrieval | Long-context English documents and RAG pipelines (8192-token context window). |
bge-m3:latest | ~1.20 GB | Multilingual | Turkish and other non-English content; 100+ language support. |
mxbai-embed-large:latest | ~0.7 GB | English & Retrieval | Search and retrieval scenarios; competes closely with hosted OpenAI embedding models. |
snowflake-arctic-embed:latest | ~0.7 GB | English & Retrieval | Enterprise-grade, data-heavy applications where retrieval precision matters. |
paraphrase-multilingual:latest | ~0.6 GB | Multilingual | Cross-language text similarity and clustering, built on Sentence-Transformers. |
all-minilm:latest | ~0.05 GB | Lightweight / Edge | Low-resource machines, fast prototyping, and basic text similarity checks. |
Note on sizes
Sizes shown are approximate, rounded download sizes for each model's default Ollama build (typically Q4 or FP16 quantization) — actual disk usage can vary slightly between Ollama versions.
Choosing a Model by Category
English & Retrieval
nomic-embed-text, mxbai-embed-large, snowflake-arctic-embed — strong general-purpose and search-oriented performance on English content.
Multilingual
bge-m3, paraphrase-multilingual — best choice when your knowledge base mixes Turkish, English, or other languages.
Lightweight / Edge
all-minilm — a 46 MB model for CPU-only machines, fast prototyping, or basic similarity checks where top-tier accuracy isn't required.
Downloading a Model
Alternatively, you can pull any model directly via the Ollama CLI on your machine (e.g. ollama pull nomic-embed-text) — it will then appear automatically in the Settings dropdown alongside the built-in download list.
Hosted (Cloud LLM) Embedding
Besides local Ollama, the Embedding sub-tab under Settings → Agentic lets you set the provider to Cloud LLM (Hosted API). Instead of Ollama, your text is turned into vectors by a hosted model through the Modelion gateway or OpenRouter, with Modelion models listed first and OpenRouter's embedding models below them.
- Where the models come from: the lists are read live from each provider, so new embedding models appear without an app update. OpenRouter models that cost nothing carry a green free badge, and a free models only checkbox narrows the list.
- No download, no local server: nothing has to be installed or running on your machine.
- Your text leaves your machine: unlike Ollama, the text being embedded is sent through TotalApp's server to the provider you chose. Keep sensitive material on Ollama if that matters to you.
Changing the model makes stored vectors outdated
Vectors produced by different models cannot be compared with each other. Every knowledge source remembers which model produced its vector; after you change the embedding model (Ollama to hosted, hosted to Ollama, or one model to another), the old vectors are recomputed automatically the next time those sources are ranked. Sources saved before TotalApp recorded this are recomputed once as well. With a hosted model this means each source's text is sent to the provider again.
FAQ
bge-m3 is the strongest choice for Turkish and other non-English content, given its multilingual training and large context support.bge-m3 are more accurate but slower to embed each source; smaller models like all-minilm run much faster but may rank less precisely. Choose based on your knowledge base size and hardware.