Local AI is unsupported. Use cloud-assisted or non-AI PolicyVector workflows.
PolicyVector Local AI
Will Local AI run well on your machine?
PolicyVector runs Local AI through Ollama on your machine. Start with memory, then use CPU and GPU guidance to choose the Qwen profile that will feel practical for real study work.
Choose your starting point.
Memory is the clearest first filter. These are orientation points; the detailed Windows and Mac tables below account for CPU, GPU, and expected workload.
Start with qwen3:4b for comfortable entry-level Local AI.
qwen3:8b is the best default; stronger machines can move toward qwen3:14b.
Pick the profile that fits the whole machine.
System RAM is the hard floor. CPU and GPU determine how responsive local generation feels, especially when PolicyVector is also indexing PDFs, building prompts, or generating study material.
| Tier | Model | RAM | CPU | GPU / VRAM | Expected Experience |
|---|---|---|---|---|---|
| Unsupported | None | Under 8 GB | Any | Any | Do not offer Local AI. Use cloud-only or non-AI workflows. |
| Basic Quality Minimum | qwen3:4b | 8 GB | 4 modern cores | GPU not required | Runs Local AI, but slower. Best for Q&A, title inference, small prompts, and light study generation. |
| Basic Quality Comfortable | qwen3:4b | 16 GB | 6+ modern cores | Optional NVIDIA 6 GB+ VRAM | Comfortable entry-level Local AI with headroom for PolicyVector, Ollama, PDFs, search, and the desktop shell. |
| Medium Quality | qwen3:8b | 32 GB | 8+ modern cores | NVIDIA 8 GB+ VRAM; 12 GB nicer | Best default for most PolicyVector users. Good for Q&A, safe-copy chunks, and normal Multiple Choice generation. |
| High Quality | qwen3:14b | 32 GB | 10-12+ modern cores | NVIDIA 12 GB+ VRAM | Good quality tier for heavier generation and better answers. CPU-only can work, but should be treated as slow. |
| Heavy Local Workstation | qwen3:14b | 64 GB+ | 12+ modern cores | NVIDIA 16 GB+ VRAM | Best for large documents, repeated generation, safe-copy work, and local Multiple Choice without pinning the PC. |
macOS guidance by Mac model.
The macOS app is built for Apple Silicon. Because Apple has a contained Mac lineup, the simplest way to choose a Local AI tier is by Mac family and unified memory.
| Tier | Mac | Memory | Suggested Profile | Expected Experience |
|---|---|---|---|---|
| Unsupported | Intel Mac | Any | Use cloud or non-AI workflows | The current macOS app is Apple Silicon/arm64. Intel Macs are not the target for this Local AI path. |
| Basic Quality Minimum | MacBook / entry MacBook | 8 GB models are not recommended for Local AI | qwen3:4b only when memory allows | Good for reading, organizing, and study work. Treat Local AI as limited unless the machine has 16 GB+ unified memory. |
| Basic Quality Comfortable | MacBook Air | 16 GB+ unified memory | qwen3:4b; qwen3:8b on newer/high-memory Airs | Best portable entry point. Works well for normal Q&A and study generation when other heavy apps are closed. |
| Medium Quality | MacBook Pro, base M chip | 16-24 GB+ unified memory | qwen3:8b | Stronger default laptop choice for Local AI, especially when PolicyVector is also indexing PDFs or building longer prompts. |
| High Quality | MacBook Pro, Pro or Max chip | 24-48 GB+ unified memory | qwen3:8b or qwen3:14b | Best laptop fit for heavier generation, repeated Multiple Choice work, and larger document sessions. |
| Desktop Medium Quality | iMac or Mac mini | 16-32 GB+ unified memory | qwen3:4b or qwen3:8b | Good desk setup for Local AI. Choose more memory if this Mac will handle large PDFs or long study-generation sessions. |
| Heavy Local Workstation | Mac mini Pro or Mac Studio | 32-64 GB+ unified memory | qwen3:14b | Best Mac desktop fit for large documents, heavier generation, and sustained Local AI work. |
How PolicyVector uses hardware.
PolicyVector, Python, and Flask are not the heavy part. Ollama and the selected model usually decide the bottleneck.
RAM
RAM decides whether the model and context have enough headroom to run without paging, instability, or failed generations.
CPU
CPU keeps fallback generation, prompt prep, PDF work, and search responsive when the app is doing real study work.
GPU / VRAM
GPU / VRAM accelerates generation when Ollama can use it. It helps speed, but it does not replace system RAM.