FAQ

Is my visitors’ data sent anywhere? In local mode (the default on capable browsers), no — page content, questions, and answers are processed entirely in the visitor’s browser and never sent to InferKit or any LLM provider. The API is contacted only once at startup to validate your key. Remote mode (paid fallback) sends the page-grounded prompt to the configured provider. See Security & Privacy.

What browsers support local mode? Any browser with WebGPU — Chrome and Edge today; Safari is rolling it out. Visitors without WebGPU automatically use remote mode on paid tiers (or see an upgrade prompt on Free).

How big is the model download? The default (small, Qwen2.5-1.5B) is ~1 GB on the first local load, then cached in the browser for instant subsequent loads. For a faster first load choose nano (Llama-3.2-1B, ~720 MB) — still grounded, just lighter. Models below ~1B aren’t offered as size tiers because they answer poorly from page context.

Will it answer off-topic questions? A deterministic grounding gate refuses questions that don’t overlap the page’s content before calling the model. Tighten it with groundingMinMatches. (Smaller local models adhere less strictly once a question passes the gate; larger models follow the on-page constraint better.)

Can I use my own LLM provider / API key? Yes — BYOK on Pro/Enterprise. Store your provider key encrypted in our vault, or point InferKit at your own endpoint so we never hold the key. Providers: OpenAI, Groq, Anthropic, Google Gemini, or any OpenAI-compatible endpoint.

How do I stop my publishable key from being abused? Publishable keys are public by design but fenced by an Origin/domain allowlist, bot protection (Turnstile), rate limits, hard quotas, and anomaly auto-suspend. For server-side calls, use a secret key with an IP allowlist.

Can I remove the “Powered by InferKit” badge? Yes, on Pro and above.

How does billing work? Self-serve via Stripe from Dashboard → Billing — upgrade/downgrade/cancel anytime. Your plan applies to all keys in your organization automatically.

Something’s broken — how do I get help? Errors include a request_id. Send it to support@synapticortex.dev and we can pinpoint the exact request.