View Source

FrameworX 10.1.5 ships configured to talk to a local Ollama with qwen2.5:7b-instruct by default. This page documents Install-LocalAI.ps1 — the idempotent setup script that gets a fresh machine into that state.

AI Integration → Local AI → First Install Walkthrough (10.1.5 draft)

10.1.5 draft. This page describes functionality scheduled for the FrameworX 10.1.5 release (~2026-04-30). When 10.1.5 ships, the "(10.1.5 draft)" suffix will be removed and this content will become the canonical install reference.

Quick start

Run Install-LocalAI.ps1 from the FrameworX AISetup folder. The script is idempotent: re-running on an already-set-up machine is safe and finishes in a few seconds.

powershell -ExecutionPolicy Bypass -File "<FX-install>\AISetup\Install-LocalAI.ps1"

Replace <FX-install> with your FrameworX install location, typically C:\Program Files\Tatsoft\FrameworX\fx-10.

What the script does

Step	Action	Skipped if
1	Pre-checks port 11434 for any conflicting service	port is already held by Ollama
2	Installs Ollama via `winget` if missing	`ollama.exe` already on disk
3	Starts `ollama serve` if not already responding	endpoint at `http://localhost:11434` already responds
4	Pulls `qwen2.5:7b-instruct` (~4.7 GB) if missing	model already in `~/.ollama/models/`
5	Smoke-tests inference via `/v1/chat/completions`	always runs (proves end-to-end working state)

When all five steps already pass, the script does nothing destructive and returns in seconds.

What to expect on first install

Item	Value
Total disk usage	~6.5 GB (1.8 GB Ollama runtime in `%LOCALAPPDATA%\Programs\Ollama\` + 4.36 GB model in `%USERPROFILE%\.ollama\`)
First-run time	~5 minutes on a 50 MB/s connection (1.8 GB Ollama installer + 4.7 GB model pull). Slower connections scale linearly.
Permissions required	None. Ollama installs per-user — no UAC / admin elevation needed.
First chat latency	~15 seconds. The model loads from disk into RAM on first call after startup or after the keep-alive window expires.
Subsequent chat latency	~500 milliseconds on a typical CPU; faster with a GPU.
Keep-alive	Default 5 minutes. Idle longer than that and the next call pays cold-load again. Set `OLLAMA_KEEP_ALIVE=24h` in the environment to keep the model resident.

Sample run output

A green run on an already-set-up machine looks like this:

FrameworX Local AI - Install and Verify
Default model: qwen2.5:7b-instruct
Endpoint:      http://localhost:11434/v1/chat/completions

==> Checking port 11434
  ok  Ollama already serving on 11434 (PID 6224)
==> Checking Ollama install
  ok  Ollama present at C:\Users\<user>\AppData\Local\Programs\Ollama\ollama.exe
==> Checking Ollama server is responding
  ok  Ollama server responding on http://localhost:11434
==> Checking model 'qwen2.5:7b-instruct'
  ok  Model present (4.36 GB on disk)
==> Smoke-testing inference (this also primes the model into RAM)
  ok  Inference returned in 10.3s: 'pong'

All green. Total time: 12.9s.
FrameworX is ready to talk to Local AI at http://localhost:11434/v1/chat/completions

Each ok line is a state-check that found things already correct and skipped the work. On a fresh machine, those ok lines are replaced with progress messages from winget install and ollama pull.

If the script reports a port conflict

If port 11434 is already held by a different process (LM Studio, llama.cpp server, oobabooga, an old test server, etc.), the script aborts with a clear message naming the offending process. To resolve:

Stop the conflicting service, or move it to a different port.
Re-run Install-LocalAI.ps1.

The script intentionally does NOT kill foreign processes on its own — port 11434 is heavily used by the LLM ecosystem and a silent process kill is the wrong default.

Running the model on a different host

By default Ollama binds localhost only. To run Ollama on a separate machine (typically a GPU server) and have FrameworX talk to it over the network:

On the Ollama host: set OLLAMA_HOST=0.0.0.0:11434 in the system environment, then restart Ollama.
Open inbound TCP 11434 in the Ollama host's firewall.
In FrameworX, edit SolutionSettings.ModelSettings to point URL at http://<ollama-host-ip>:11434/v1/chat/completions.

A future revision of this script will accept a -RemoteHost flag to automate steps 1 and 2.

Choosing a different model

The default qwen2.5:7b-instruct is a balance of quality and footprint for typical SCADA hardware. To use a different model:

Pull it with Ollama: ollama pull <model-name> (for example, qwen2.5:3b for low-RAM gateways, qwen2.5:14b-instruct for capable servers, llama3.1, mistral, etc.).
In FrameworX, edit SolutionSettings.ModelSettings and set the Name field to the new model name.

Any OpenAI-compatible endpoint works — including cloud LLMs (OpenAI, Azure OpenAI, Anthropic via OpenAI-compat proxy). Set URL and Authorization accordingly.