QUESTIONS, answered
BEFORE YOU ASK.

Today: public website URLs (HTTP/HTTPS), text-based PDFs uploaded to us, and plain text you paste in. We can't ingest scanned PDFs (those need OCR — coming later) or pages that require a password to view. Sitemap-based whole-site crawl is also coming later; for now, add the important URLs one at a time.

Most single URLs and short PDFs finish in 10–60 seconds. Large PDFs (50+ pages) or long pages take longer because each chunk has to be embedded by OpenAI. If a source sticks at pending for more than 5 minutes, refresh — it's probably already done. If it shows error, the message in the status tells you what went wrong (usually: blocked by the target site, or a scanned PDF).

Yes — that's the normal case. Free is 3 sources/bot, Solo 10, Builder 25. The bot retrieves from all sources at once, so adding more good material almost always improves answers; adding noisy or contradictory material can hurt them.

Not yet. Our ingest only fetches what an unauthenticated visitor would see, with a 10-second timeout and up to 5 redirect hops. If your content lives behind login, the workarounds today are: export the content to a public URL, paste the raw text, or upload it as a PDF. Auth-protected crawl is a Phase 5 item — not committed.

Three usual reasons. One: the information isn't actually in your sources — add the page or paste the text that answers the question. Two: the relevant chunk doesn't pass our similarity threshold for that exact phrasing — try asking the question the way a real visitor would, and consider rewording your source content to use the same words. Three: the bot is correctly saying "I don't know" because the answer genuinely isn't there — that's the bot doing its job (better than hallucinating).

Every answer ships with the source chunks that informed it. We retrieve the top 5 most similar chunks (cosine similarity over OpenAI's 1536-dim embeddings, filtered to your bot), and the answer cites which ones it used. If a question can't be answered from your sources at all, the bot says so explicitly instead of making something up — that fallback is non-negotiable.

Not via a "preferred answer" override yet. The bot composes its reply from your sources at query time — so the way to change what it says is to change what it can read. Add a clearer FAQ page, paste a "house style" note, or remove a source that's confusing it. Direct answer-editing is on the roadmap.

Your bot runs on frontier Claude models from Anthropic — the same family that tops the reasoning benchmarks. We pick the right model for each tier and upgrade it automatically as Anthropic ships better ones, so you never have to manage model versions. Retrieval embeddings are handled by OpenAI. (Opus-class reasoning is on the 2026 roadmap.)

Free ($0): 1 site, 3 sources, 100 messages/mo — live on your real site, with guided page actions included. Solo ($9/mo): 1 site, 10 sources, 1,000 messages/mo, plus lead capture. Builder ($24/mo): 2 sites, 25 sources/site, 3,500 messages/mo, plus intent insights and Saavos branding removed. Yearly billing on either paid tier is 10× monthly (about 17% off).

Yes, anytime, from the customer portal (Dashboard → Billing → Manage). Cancellation takes effect at the end of the current billing period — you keep access until then, no penalty, no contract. Your data stays accessible during that grace window.

There is no overage on any plan. Every tier just pauses at its cap and resumes next cycle — no per-message charges, no surprise bills, ever. We'd rather you upgrade on purpose than get an invoice you didn't expect.

On every plan — Free, Solo, and Builder — the bot shows visitors a friendly "limit reached" note and resumes at the start of your next billing cycle. No per-message overage, no surprise charges. Usage resets at the start of each Dodo billing period. If you keep hitting the cap, that is the signal to move up a tier.

If you upgrade and decide within 7 days that saavos is not right for you, email support@5minbot.com and we will refund the charge in full — no questions asked. After 7 days, cancellation stops future charges but past periods are not refunded. The no-card preview exists precisely so you do not need to gamble money to evaluate.

Copy one <script> tag from the Install tab in the dashboard, paste it before </body> on your site. Works on any HTML site — WordPress, Webflow, Framer, Shopify, plain HTML. The widget loads asynchronously, so it doesn't slow your page; the chat itself opens in an iframe at /embed/<your-bot> only when a visitor clicks the launcher.

Yes. The widget is responsive — on small screens it opens full-bleed so visitors aren't squeezed into a corner. Tested on current iOS Safari, Android Chrome, and the major in-app browsers (Instagram, LinkedIn). If you find a host site or browser where it misbehaves, send us the URL and we'll fix it.

Limited customization today: a primary color and the launcher position (bottom-right or bottom-left) on the dashboard's Customize tab. Logo, full theme tokens, and CSS overrides are coming later. If you need pixel-level control before that ships, the embed iframe is just a URL — you can build your own launcher and iframe it yourself.

saavos is AI-only by design — we don't run a live-chat product. When the bot can't confidently answer (or a visitor explicitly asks for a person), the fallback routes them where you configure it: a support email, a calendar link, a help-desk URL, or a custom "we'll get back within X hours" message. The intent is that you don't need to sit in a chat queue.

Conversation data is tenant-isolated at the database level — we use Postgres row-level security so one bot's chats are physically unreachable from another bot's queries, not just hidden by app code. We don't sell, share, or train models on your data, and Anthropic and OpenAI's no-train enterprise terms apply to all calls we make on your behalf.

Application data (bots, sources, chunks, conversations): Supabase Postgres, currently in the AWS us-east-1 region. File uploads (PDFs): Supabase Storage in the same region. Embeddings are sent to OpenAI and answers to Anthropic in real time; neither provider stores prompts or completions for training under the enterprise terms we operate under.

Yes for the requirements we control. We store the minimum data needed to run the bot (auth email, bot config, sources you uploaded, conversation logs). You can export and delete everything from the dashboard. We're not a Data Processor in a contractual sense yet (no signed DPA template) — if you need a DPA, email founder@5minbot.com and we'll send one.