docs/LLM_STORAGE_AND_DOWNLOAD_POLICY.md

# LLM storage and download policy

## Where the LLMs live

Built key:

```text
models/ollama/   Ollama model blobs
models/gguf/     optional llama.cpp GGUF files
models/stt/      speech-to-text models
models/tts/      text-to-speech models
models/vision/   vision models or pointers
```

## Customer insertion flow

Normal use does not download models.

```text
Insert key
open START-HERE.cmd
runtime starts
models are read locally from the key/cache
```

## When downloads happen

Downloads happen only in these cases:

1. factory preloading before shipping,
2. SaaS-generated package construction,
3. explicit model add action,
4. signed update/model pack.

## Ollama storage

Windows launcher sets:

```powershell
$env:OLLAMA_MODELS = "$DriveRoot\models\ollama"
```

Linux services set:

```bash
OLLAMA_MODELS=/opt/jackailocal/models/ollama
```

This makes Ollama read/write models inside the JackAILocal key/runtime path.

## Why not download at insertion

Downloading at insertion breaks the product claim:

- no offline mode,
- no airplane mode,
- no guaranteed first-run experience,
- customer sees model management complexity,
- support load explodes.

Correct model:

```text
Build once -> use offline repeatedly.
```