AI Features & Privacy

Overview

AssisT uses a privacy-first hybrid AI system that gives you four ways to use AI features, from completely offline to cloud-powered. Your data stays on your device by default. All 8 AI features are routed through a shared AI client that consistently handles mode detection, availability checking, and response generation.

Two settings shape what comes back from every one of them. Personalised AI feeds your learning profile into each AI feature rather than only one or two, so the output matches how you said you read and work. The reading-level dial sets how simply summaries and simplifications are written, so you can ask for the same passage again at a level that suits you. Both are companion preview features and can be switched off individually.

Four AI Modes

AssisT offers flexible AI processing with four distinct modes:

Mode	Privacy	Cost	Performance	Requirements
Off	N/A	Free	Features disabled	None
Local AI (Ollama)	100% Private	Free	Good (hardware-dependent)	Ollama installed
Browser AI (WebLLM)	100% Private	Free	Good (GPU-dependent)	WebGPU-capable browser
Cloud AI (API)	Your API only	Free (Gemini) or paid	Excellent	API key

You can switch between modes instantly using the inline AI mode switcher in the popup — click Off, Cloud, Browser AI, or Local AI chips to switch without leaving the page.

Key Principles

• Your Choice: Pick the AI mode that matches your privacy and performance needs
• No Data Collection: We never see, store, or transmit your data
• Bring Your Own Key: Cloud mode uses your own API keys, not ours
• Graceful Fallback: Features work even without AI (with reduced functionality)

Local AI with Ollama

AssisT integrates with Ollama, a free, open-source tool that runs AI models directly on your computer.

Why Local AI?

Benefit	Description
Privacy	Data never leaves your device
Compliance	No text leaves the device, so nothing is disclosed to any provider
No Cost	No API fees or subscriptions
Offline	Works without internet connection
Speed	No network latency for requests

Live Model Selector

AssisT auto-detects all models installed on your Ollama instance and shows them in a live dropdown in the popup’s Local AI panel. You can:

• See all your installed models at a glance
• Switch between models without restarting
• Your selection persists across sessions

Recommended Models

Model	Size	Best For
qwen3:8b-q4_K_M	5GB	Default — best JSON compliance, instruction following
llama3.1:8b	5GB	Strong general-purpose, good reasoning
gemma3:4b	3GB	Fast responses for basic tasks
deepseek-r1:8b	5GB	Code and reasoning tasks
mistral:7b	4GB	Complex analysis, detailed responses
llava	4GB	Image understanding (vision)

Task-Optimised Model Routing

AssisT automatically selects the best available model for each task:

Task	Priority Models
Knowledge Graph	gemma3:4b, qwen3:8b
Socratic Tutor	qwen3:8b, llama3.1:8b
Citation Analyzer	qwen3:8b, mistral:7b
Study Path Generator	qwen3:8b, llama3.1:8b
Assignment Breakdown	qwen3:8b, llama3.1:8b
Summarisation	Any available model
Text Simplification	Any available model

Reasoning Models

qwen3 and deepseek-r1 think out loud before answering. AssisT handles that: their thinking text is kept out of the answer you see, and requests that need structured JSON back (Knowledge Graph, Citation Analyzer, mind-map building) still get valid JSON from them.

Permission

AssisT does not ask for access to your local AI server at install time. The loopback address it needs (localhost or 127.0.0.1, port 11434) is an optional permission, requested only at the moment you actually turn local AI on.

So the first time you switch Local AI on, or press Apply on a custom server address, Chrome shows an “Allow” prompt. Choose Allow and AssisT can reach Ollama. If you dismiss it, or press Check outside that prompt, the popup tells you that local AI needs your permission rather than telling you to start Ollama, because a missing permission and a stopped server look identical from the outside and only one of them is fixed by starting Ollama.

Two things follow from this. A student who never uses local AI is never asked for the permission at all, which is the least-privilege position this project holds itself to. And because Chrome will only show the prompt during a press, it appears on the switch or on Apply, never on its own later.

You can withdraw the permission whenever you like, under Chrome’s extension settings for AssisT. Local AI then stops working until you grant it again, and everything else is unaffected.

Server Address

By default AssisT looks for Ollama at the usual loopback address. If your install listens on a different port, set the address in the Local AI panel and AssisT will use it. The public build only accepts loopback addresses, because a custom remote host is blocked by the extension’s content security policy by design; an institution running a shared Ollama server for its students adds that host to its own build.

Installing Ollama

1. Download Ollama from ollama.com/download
2. Install and run Ollama on your computer
3. Pull a model: ollama pull qwen3:8b-q4_K_M
4. AssisT will automatically detect it

How Local AI Works

Your Browser (AssisT)
        ↓
    Message Bridge
        ↓
Ollama (localhost:11434)
        ↓
    AI Response
        ↓
Back to AssisT

All communication happens locally on your machine. Nothing is sent to external servers.

Browser AI with WebLLM

AssisT supports browser-native AI powered by WebLLM and WebGPU. Models run entirely inside your browser tab — no server, no API key, no software to install.

Why Browser AI?

Benefit	Description
Zero Install	No Ollama or other software needed
Privacy	All processing happens in your browser
No Cost	Completely free
Portable	Works on any WebGPU-capable device

Available Models

Model	Size	Best For
llama-3.2-1b	~1GB	Ultra-fast, basic tasks
gemma-2b	~2GB	Balanced small model
phi-3.5-mini	~3GB	Strong reasoning for its size
qwen2.5-3b	~3GB	Good multilingual support
llama-3.2-3b	~3GB	Capable general model
mistral-7b	~5GB	Complex analysis
llama-3.1-8b	~6GB	Most capable browser model
gemma-9b	~5.4GB	Gemma 2 9B, strongest reasoning

Model Management

The popup shows three model states:

• Loaded (green) — Ready to use, currently in memory
• Cached (blue) — Downloaded to device, loads quickly
• Available (grey) — Needs downloading first

Models are downloaded once and cached in browser storage. Subsequent loads are fast.

How Browser AI Works

Your Browser (AssisT)
        ↓
    WebLLM Engine (WebGPU)
        ↓
    GPU-Accelerated Inference
        ↓
    AI Response (in-browser)

Everything happens inside your browser. No external connections whatsoever.

Requirements

• Chrome 113+ or any browser with WebGPU support
• Dedicated GPU recommended for larger models (integrated GPUs work for 1-3B models)
• Sufficient storage for model downloads (1-6GB per model)

Cloud Providers (Optional)

For users who want more powerful AI capabilities, AssisT supports multiple cloud providers through API keys you provide.

Supported Providers

Provider	Strengths	Best For
Anthropic (Claude)	Coding, academic writing, analysis	Text simplification, tutoring
OpenAI (ChatGPT)	Creative, conversational	Brainstorming, general tasks
Google (Gemini)	Multimodal, visual, factual	Image understanding
Perplexity	Real-time web, citations	Research, fact-checking

Bringing Your Own API Key

1. Get an API key from your preferred provider:
   • Anthropic Console (Claude)
   • OpenAI Platform (ChatGPT)
   • Google AI Studio (Gemini)
   • Perplexity Settings
2. Open AssisT settings
3. Go to AI Settings > Cloud Providers
4. Select your provider and enter your API key
5. Choose your preferred model

Cost vs Quality

Every provider offers a fast, cheap model and a slower, more capable one, and the model picker groups them that way:

Model type	Cost	Best for
Fast	Cheapest per token	Summaries, simplification, high volume
Balanced	Moderate	Most everyday use
Quality	Highest per token	Knowledge Graph, Multi-Document Compare, anything where accuracy matters

Tip: Start with the fast model. Move up only when an answer is not good enough.

API Key Security

• Your API keys are encrypted with AES-256 and stored locally in Chrome’s secure storage
• They are never sent to Fiavaion servers
• Only transmitted directly to the provider when you use cloud features
• You can remove them anytime from settings

Which models you can choose

AssisT does not ship its own list of model names for you to memorise. When you pick a provider, the model dropdown shows the current models that provider offers, with a recommended default already selected. Model names change several times a year, so they are not printed in this documentation; open the dropdown to see what is available today.

Provider	Cost	Notes
Google Gemini	Free tier, no credit card	The default provider for anyone starting out. Gemini 2.5 Flash is the default model. Daily limits reset automatically, and you are never charged unless you deliberately set up billing on Google Cloud
Anthropic	Paid	Claude models. Strong on careful reading and explanation
OpenAI	Paid	GPT models. You need credit on the account first
Perplexity	Paid	Sonar models, unique in providing real-time web access with citations, so they suit research and fact-checking

Recommendation: If you have no API key and no strong preference, use Google Gemini. It is free, the key takes about a minute to create, and the free daily allowance covers normal student use.

Feature defaults: each AI feature asks for a model suited to it rather than always using the same one, so a summary is answered by a fast model while Knowledge Graph and Multi-Document Compare go to a more capable one. You can override the choice in AI settings.

Gemini Nano (Experimental)

Chrome’s built-in Gemini Nano model provides on-device AI processing without installing anything.

Status

Gemini Nano support is experimental and takes second place to WebLLM (Browser AI), which offers more models, better control and wider browser compatibility. It remains available for advanced users who have already enabled Chrome’s experimental flags. Detection was fixed in 1.0.0: if you have the flag on and the model downloaded, AssisT now finds it instead of reporting it as unavailable.

Requirements

1. Chrome 128 or later (Canary, Dev, Beta, or Stable)
2. Feature flag enabled: Visit chrome://flags/#optimization-guide-on-device-model and set to “Enabled”
3. Model download: Chrome downloads the model automatically on first use

Gemini Nano vs WebLLM vs Ollama

Feature	WebLLM (Browser AI)	Gemini Nano	Ollama
Setup	None — just download a model	Chrome flag required	Install separate app
Model Choice	8 models (1B-8B)	Single model (Google’s)	Unlimited models
Performance	Good, GPU-accelerated	Basic tasks only	Best for complex tasks
Browser Support	Chrome 113+	Chrome 128+ with flags	Any browser
Customisation	Choose model per task	Limited	Full control

Recommendation: Use Browser AI (WebLLM) for zero-install private AI, or Ollama for maximum capability. Gemini Nano is best suited for users already familiar with Chrome feature flags.

How the AI Mode System Works

AssisT routes all AI requests through a shared AI feature client (ai-feature-client.js) that provides consistent mode detection, availability checking, and response generation across all 10 features.

Feature Request (any of 10 AI features)
      ↓
Shared AI Feature Client
      ↓
   getAIMode() → reads aiMode from storage
      ↓
   ┌────────┬────────┬──────────┬─────────┐
   │        │        │          │         │
  OFF    Cloud    WebLLM     Ollama    Gemini
   │        │        │          │      (experimental)
   │    API Key   WebGPU   localhost
   │    + Model   Engine   :11434
   │        │        │          │
   └→ Status  Generate  Generate  Generate
      bar       ↓        ↓        ↓
              Response  Response  Response

Persistent Status Bars

Every AI feature displays a persistent status bar showing:

• Orange warning when AI is unavailable (with setup links)
• Green success after a successful generation
• Mode indicator showing which AI backend processed the request

Feature Compatibility by Mode

Feature	WebLLM	Ollama	Cloud
Summarisation	✅	✅	✅
Text Simplification	✅	✅	✅
Assignment Breakdown	✅	✅	✅
Socratic Tutor	⚠️ Basic	✅	✅
Multi-Doc Compare	⚠️ Basic	✅	✅
Knowledge Graph	⚠️ Small texts	✅	✅
Citation Analyzer	✅	✅	✅
Study Path Generator	⚠️ Basic	✅	✅
Image Understanding	❌	✅ (llava)	✅
Research & Citations	❌	❌	✅ (Perplexity)

Fallback Behaviours

When AI isn’t available, features gracefully degrade:

Feature	Fallback Behaviour
Summarise	Shows first paragraph
Simplify	Feature disabled with status message
Image Describe	Requires vision model
Knowledge Graph	Disabled with status message

Privacy Guarantees

What We Never Do

• Collect or store your data
• Send data to our servers
• Track your AI usage
• Share information with third parties

What Stays Local

• All text you process
• Documents you summarise
• Images you analyse
• Conversation history

What this means for GDPR and FERPA

With Local AI or Browser AI, nothing you process is transmitted anywhere:

• GDPR: Fiavaion processes no personal data. With a cloud provider, that provider is your own sub-processor under your agreement with them, not a sub-processor of Fiavaion’s
• FERPA: AssisT is designed with FERPA principles in mind — no student data is collected
• HIPAA: no claim is made. AssisT is not a healthcare application and has not been assessed against HIPAA

Institutions evaluating AssisT should read the privacy and data-processing statement and the sub-processor list.

Performance Tips

For Best Local AI Performance

1. Use an SSD: Faster model loading
2. 8GB+ RAM/VRAM: Required for larger models
3. Keep Ollama Running: Faster first response
4. Choose Appropriate Models: Match model size to your hardware

Why Memory Matters

• More VRAM = Better Models: With more video memory (or unified memory on Apple Silicon), you can run larger, more capable models
• More Memory = Longer Context: Additional memory allows longer context windows—the AI can “remember” more of your document
• Longer Context = Fewer Hallucinations: When AI sees more context, it makes fewer mistakes because it has more information to work with

Memory Types

Type	What Matters	Notes
Dedicated GPU	VRAM (8GB good, 12GB+ great)	NVIDIA/AMD graphics cards
Apple Silicon	Unified memory (16GB good, 32GB+ excellent)	M1/M2/M3/M4 Macs
CPU-only	System RAM (16GB min, 32GB recommended)	Slower but works

Recommended System Requirements

Setup	RAM/VRAM	Storage	Models
Minimal	8GB	4GB free	gemma3:4b
Standard	16GB	8GB free	qwen3:8b-q4_K_M
Full	32GB+	15GB free	Multiple models + longer context

Troubleshooting

Ollama Not Detected

1. Ensure Ollama is installed and running
2. Check that it’s accessible at localhost:11434
3. Open a browser and visit localhost:11434 — you should see “Ollama is running”
4. Restart Ollama if needed
5. Refresh the AssisT extension

Ollama Model Not Showing in Dropdown

1. Verify the model is installed: run ollama list in your terminal
2. Ensure Ollama is running (the dropdown fetches models live)
3. Try closing and reopening the AssisT popup

WebLLM Model Won’t Load

1. Check your browser supports WebGPU — visit chrome://gpu and look for “WebGPU” in the feature list
2. Ensure you have enough GPU memory for the model (1-6GB depending on model)
3. Try a smaller model (llama-3.2-1b or gemma-2b)
4. Close other GPU-intensive tabs or applications
5. Restart Chrome if the GPU context is corrupted

WebLLM Shows “Not Downloaded”

1. Click the Download button next to the model
2. Wait for the download to complete (shows progress bar)
3. Larger models (7B+) may take several minutes depending on connection
4. Downloaded models are cached in browser storage — they persist across sessions

Slow AI Responses

1. Local AI: Try a smaller Ollama model (gemma3:4b is fastest)
2. Browser AI: Use a smaller WebLLM model (1B-3B range)
3. Ensure no other AI requests are processing simultaneously
4. Check your system’s available memory/VRAM
5. Close other resource-intensive applications

AI Feature Shows Orange “Unavailable” Bar

This status bar appears when the selected AI mode can’t be reached:

• Cloud mode: Check your API key is entered and valid
• Local mode: Ensure Ollama is running at localhost:11434
• Browser AI: Load a WebLLM model first (click the status bar for setup)
• Off mode: Switch to an active AI mode using the mode chips in the popup

AI Response Times Out

Some AI operations (especially Knowledge Graph and Multi-Doc Compare) may timeout on slower hardware. AssisT uses a 25-second timeout for all async operations. If you experience timeouts:

1. Try shorter input text
2. Use a faster model
3. Switch to Cloud AI for complex operations

Model Download Failed

1. Check your internet connection
2. Ensure enough disk space is available
3. Try downloading a smaller model first
4. Restart Ollama (for local) or Chrome (for WebLLM) and try again