OpenAI, Mistral, Gemini, Albert
Cloud APIs reached with an API key declared in the environment of the server (see the Show Variables dialog of r-ioga). Unchanged: pick the provider, then the model.
Local AI service (Ollama)
Run the AI assistants of r-io suite without API key nor quota: Ollama embedded in the server, installed and started on demand with a model recommended for the hardware, or an Ollama server already running, reached by its URL.
01 · Two families
External providers or a local service
The AI assistants (model invention, RAG, picture analysis, command assistant) use the AI service chosen on each project in r-ioga Edit project. The choice is split into two families.
Ollama, embedded
The server downloads Ollama itself (native binary, no Docker), starts it on demand on its own port and pulls the chosen model. Nothing leaves the machine; no key, no quota.
Ollama, external
An Ollama server already running on the network (for example a GPU server of the lab). Only its URL is entered; it is stored in the project.
02 · In Edit project
Choose the local service
In the IA & RAG Knowledge documents card, the first list, IA Service, switches between External provider and Local service. With the local service, the second list becomes the Ollama mode (Embedded or External) and a panel shows the state of the service read on the server.
The hardware is the one of the server
Ollama embedded runs on the machine that runs r-io suite, not on the browser machine. The panel therefore shows the memory and the GPU of the server, and the recommendation depends on them.
03 · Embedded mode
What the server does by itself
OllamaEmbeddedRuntime manages one Ollama per machine, shared by every distribution of the machine (like the embedded MongoDB binaries).
# 1. find a binary ~/.embeddedOllamaRio/bin/ollama(.exe) # installed by r-io suite or an Ollama already installed # PATH, %LOCALAPPDATA%\Programs\Ollama, /Applications/Ollama.app # 2. otherwise download the release archive (about 1.4 GB on Windows) github.com/ollama/ollama/releases/latest/download/ollama-windows-amd64.zip | ollama-linux-amd64.tar.zst | ollama-darwin.tgz # 3. start on demand (first AI call or installation) OLLAMA_HOST = 127.0.0.1:11441 # system property rio.ollama.port OLLAMA_MODELS = ~/.embeddedOllamaRio/models ollama serve >> ~/.embeddedOllamaRio/ollama.log
Never the default Ollama port
The embedded server listens on 11441, not 11434, so it never collides with an Ollama started by the user. A server already answering on 11441 (started by another instance of the machine) is reused and is not stopped by this instance.
Started on demand, stopped with the server
Nothing starts at boot: a distribution that does not use the local service pays nothing. The process started by the server is stopped with it; Ollama itself unloads an idle model after five minutes.
Downloaded once
Binary and models live in the user home and are shared: a model of several gigabytes is downloaded once for every distribution and every worktree of the machine.
04 · Choosing a model
A catalog and a recommendation
The model list merges the r-io suite catalog (ollama/ollama-model-catalog.json) and the models already installed on the Ollama server. Icons: a star marks the recommended model, a check an installed model, a download arrow a model to install.
What is detected
OllamaHardware reads the physical memory, the cores and, with nvidia-smi, the NVIDIA GPU with the most memory. On Apple Silicon two thirds of the unified memory are counted as GPU memory. Other GPUs are not detected: the CPU rule then applies.
The biggest model that fits
With a GPU, a model fits when its size is at most 80 % of the GPU memory (the rest is kept for the context). Without GPU, at most 40 % of the RAM and 7 GB, beyond which answers take minutes. Among the models that fit, a model fitting in the GPU comes first, then the biggest, then one accepting images. When nothing fits, the smallest model.
Laptop with an 8 GB GPU
An NVIDIA RTX 2000 Ada (8.6 GB) accepts models up to about 6.9 GB: gemma4:e4b (6.58 GB, text and images) is recommended; gemma4:12b and bigger are marked too big for the server.
| Model | Size | Images | Note |
|---|---|---|---|
llama3.2:3b | 2.0 GB | — | Small and fast |
qwen3:4b | 2.5 GB | — | Small, good at structured JSON |
qwen3.5:4b | 3.3 GB | yes | Small |
gemma3:4b | 3.4 GB | yes | Small |
qwen3:8b | 5.2 GB | — | Balance for the RAG extraction |
qwen3.5:9b | 6.6 GB | yes | |
gemma4:e4b | 6.6 GB | yes | Made for laptops |
gemma4:12b | 8.0 GB | yes | |
qwen3:14b | 9.3 GB | — | |
gpt-oss:20b | 13.8 GB | — | Reasoning |
mistral-small3.2:24b | 15.2 GB | — | Same family as the Albert model |
qwen3.5:27b | 17.5 GB | yes | Big GPU |
gemma4:31b | 20.4 GB | yes | Default model of Albert, big GPU |
Sizes are the download sizes read on the Ollama registry on 2026-10-16; “images” means the manifest carries a projector layer. Once a model is installed, its real capabilities are read on the server (POST /api/show) and take precedence over the catalog.
05 · Install
Download with a progress bar
When the selected model, Ollama itself (embedded mode) or the RAG embedding model is missing, the panel shows an Install button. The installation runs in background on the server (OllamaInstallManager); r-ioga polls its progress every second without opening the loading dialog.
DOWNLOAD_OLLAMA # embedded mode, when no binary is found START_OLLAMA # embedded mode PULL_MODEL # POST /api/pull, streamed progress summed over the layers PULL_EMBEDDING_MODEL # nomic-embed-text (0.3 GB), when missing READY
- Cancel stops the current download. Ollama keeps the layers already downloaded and resumes them at the next installation.
- The project is saved only when its model is ready: Save shows a warning while a download runs or while the model is missing.
- On a laptop of the laboratory, Ollama,
gemma4:e4bandnomic-embed-textwere installed in about four minutes (around 40 MB/s).
06 · External mode
An Ollama server reached by its URL
Choose External, enter the URL of the server (default proposal http://localhost:11434) and click Connect. The list then shows the catalog and the models installed on that server. The URL is stored in the project.
- The URL is reached from the server, not from the browser:
localhostdesignates the machine of r-io suite. - When no server answers, the panel says so and proposes nothing to install.
- Install also works on an external server: the model and the embedding model are pulled on that server.
07 · At run time
How the local model is called
OllamaChatModelBuilder builds one LangChain4j chat model per server, model and settings, and reuses it.
The project settings apply
Model, temperature and maximum tokens of the context are applied. Answers are requested in JSON, the context window is 16,384 tokens (Ollama loads 4,096 by default, too short for the RAG prompts) and a request is abandoned after ten minutes. Reasoning models are called with thinking off.
Picture analysis
If the project model does not accept images, another installed model accepting images is used, with a warning in the log. If none is installed, the error asks to install one from Edit project.
Cold and warm calls
The first call loads the model into memory: about 100 s for gemma4:e4b the first time on the laboratory laptop. Then, with the model fully on the GPU, a short JSON answer came back in less than one second.
The RAG embeddings use nomic-embed-text on the same server as the chat model. Changing the provider of a project still empties its RAG store: ingest the documents again after the change.
08 · Project & API
What is stored, what is exposed
| Project property | Value | Read by |
|---|---|---|
IA Provider Name | OLLAMA | AIChatbotClient.createChatContext → ollamaContext of ai-chatbot.xsd |
IA Model Name | Ollama name with its tag, e.g. gemma4:e4b (free text) | |
IA Ollama Mode | EMBEDDED (default) or EXTERNAL | |
IA Ollama URL | URL of the external server, required in external mode |
| REST resource (r-ioga) | Role |
|---|---|
GET …/generic-application/aichatbot/ollama/status?mode=&url= | Hardware, embedded runtime, reachability, models with the recommended one |
POST …/ollama/install {mode, url, model} | Starts a background installation, returns the task |
GET …/ollama/install/{taskId} | Phase, percentage, bytes, speed, error |
POST …/ollama/install/{taskId}/cancel | Cancels the running installation |
The former keys ollama and ollama-embedded of unified_config.properties no longer exist: everything is configured per project.
09 · Limits
What to know
- The catalog models have not been validated one by one with the r-io suite prompts: a small model may return incomplete extractions.
- Only NVIDIA GPUs (and Apple Silicon) are detected; with another GPU the recommendation follows the CPU rule and stays cautious.
- When the server process is killed (not stopped),
ollama.exekeeps running on 11441; the next start reuses it. - No image creation with Ollama: the features that create pictures need an external provider.
10 · Sources
Implementation
gind-governance-ai-chatbot, packageservice.builder.model.ollama:OllamaChatModelBuilder,OllamaEmbeddedRuntime,OllamaHardware,OllamaModelCatalog,OllamaInstallManager; catalogsrc/main/resources/ollama/ollama-model-catalog.json.ai-chatbot.xsd(ollamaContext,OllamaMode) andAIChatbotClient;Project_MetaModel.xml(IA Ollama Mode,IA Ollama URL).AIChatbotResource(REST resources) and r-iogaedit-project-page,ai-chatbot.service.ts.