Frequently asked questions

It reads content you give it — your website, help docs, uploaded files, pasted text — and turns it into a chat assistant that answers your customers' questions on your site, 24/7.

Every answer comes with a link to the source it was drawn from, and a confidence score. If nothing in your content answers the question well enough, it refuses and offers to pass the visitor to a human rather than guessing.

No, and you should ask that of everything in this category right now.

A model provider gives you a completion endpoint. What we built is everything around it: an HTML extractor that keeps the answer text and discards the nav and cookie banners; sentence-aware chunking with overlap so answers are never cut in half; a BM25 search index built per customer over their own content; a relevance floor that produces the "I don't know" behaviour; citation tracking; and a content-gap report derived from the questions that fell below the floor.

The proof: with no AI model connected at all, the assistant still works — it retrieves and quotes your own words with sources. The generative layer rewrites those passages into conversational prose. Retrieval quality is the product; generation is a presentation layer.

It is built specifically not to. Retrieval runs first, and if the best match in your content does not clear a relevance threshold, the assistant does not answer at all — it says it could not find that and offers a handover.

When it does answer, the model is instructed that the retrieved passages are the complete extent of what it knows, and is told never to invent prices, policies, dates, availability or contact details. Every reply shows its sources so a customer can check.

No system is perfect. That is why low-confidence answers are visibly flagged rather than presented with the same certainty as high-confidence ones.

No. Creating an account, building an assistant and indexing your whole site are all free — but the assistant will not answer visitors until you buy a pack of answers.

We would rather be straight about why: a free allowance sitting behind a public chat widget can be drained by a script in an afternoon, and the compute bill lands on us regardless. So instead of a trial we made the smallest entry point cheap — 200 answers for EUR 24, a single payment with no subscription and nothing to cancel. If it does not work for you, you are out EUR 24 rather than tangled in a subscription.

About five minutes. Create an assistant, paste your website URL, wait for the crawl, test it in the dashboard, then paste one <script> tag into your site's HTML. No plugin, no DNS change, no redeploy needed for later changes.

A website URL (we crawl same-domain pages), uploaded .txt, .md, .csv, .json and .html files, and text pasted straight into the dashboard. Pasting is the reliable option when a site blocks crawlers or the content lives somewhere behind a login.

PDF and Word ingestion is the most-requested addition and is on the roadmap.

Yes. A visitor writing in German gets an answer in German, drawn from your English content, with no separate translation workflow. Retrieval works across languages less well than within one, so if most of your customers speak a language your content is not written in, add a page in that language.

Re-crawl whenever your site changes — it takes one click and replaces the old content. On the Scale plan you can schedule that automatically. The citations are the practical safeguard: if an answer links to a page you have since changed, you can see it immediately.

That is arguably the best part. Every conversation is logged with the confidence behind each answer, and the content-gap report clusters the questions it could not answer. "Eleven people asked about bulk pricing this week" is a content roadmap you did not have before.

Yes, on Growth and Scale. Ask questions server-to-server, list assistants, and pull the content-gap report as JSON. Keys are created and revoked from your dashboard, stored hashed, and shown in plaintext exactly once.

No, and on purpose. Everything is a one-off pack of answers — 200 for EUR 24, 1,000 for EUR 99, or 5,000 for EUR 399. You pay once, nothing renews, and there is nothing to cancel. Packs are valid for 12 months.

Bigger packs are cheaper per answer. Packs stack, and the one expiring soonest is always used first, so buying ahead never wastes anything.

They are not money. You buy a quantity of answers, not credit in euros. They are usable only with this service, cannot be transferred to anyone else, and cannot be exchanged for cash. We are deliberate about that: a euro balance you top up would be stored value, which is a different and much more heavily regulated product.

Bought by mistake and not used? Tell us within 14 days and we refund it in full. Partly used, within 30 days? We will usually refund the unused portion. Unused answers expire after 12 months, and we tell you the date before you buy.

Either. Signing up is free and open to anyone — an email and a password, nothing else.

At checkout you tell us which you are, because the two carry genuinely different rules. As an individual you get a 14-day right of withdrawal: email us within 14 days and you pay only a proportionate amount for the access you used. As a business, consumer withdrawal rights do not apply — our Refund Policy applies instead.

In EUR, once, when you buy a pack. Nothing is added on top. Your payment is made to Stripe, which acts as merchant of record and issues the receipt, so the charge appears on your statement from Stripe.

Stripe processes all payments. Card details are entered on Stripe's own hosted page and never touch our servers — we cannot see, store or retrieve them. We hold only a customer reference, the purchase history and invoice metadata. See Security.

There is nothing to cancel. A pack is a single payment — no card is stored for future charges and nothing renews. When the answers are used up, Sprosimo stops and tells visitors to contact you. If you want it offline sooner, remove the script tag from your site.

Read the Refund Policy, but in short: if it did not work as described, we refund the pack. If you bought one and used none of it, ask within 30 days and we refund it. We would rather return the money than have an unhappy customer or a payment dispute.

Email support@sprosimo.net with the date and amount before you contact your bank. We can identify the account, show you the usage behind it, and refund the same day if it should not have been made. That is faster for you than a chargeback.

Your content is indexed so the assistant can search it, and stored so you can manage and remove it. We do not train models on it, we do not sell it, and we do not use one customer's content to answer another customer's questions. Index data is scoped to a single assistant and every query is filtered by owner.

Only the retrieved passages, and only when the generative layer is enabled. Retrieval itself never leaves our systems, and with no model key configured nothing is sent to any provider at all.

Every third party that can touch your data is named on the Sub-processors page, with its role and location.

We store the question, the answer and the page it was asked from. We deliberately do not store visitors' raw IP addresses — a salted daily hash is kept instead, which is enough to count unique visitors and nothing more. Email addresses are stored only when a visitor types one in to request a human.

For that data you are the controller and we are your processor under the Data Processing Addendum. Tell your visitors about the assistant in your own privacy policy.

No, and we will say so plainly rather than implying otherwise. We are a small company. What we can give you is the Security page describing exactly what we do and do not do, a signed DPA, a named sub-processor list, and honest answers to a security questionnaire. If a certification is a hard procurement requirement, we are not the right vendor yet.

Yes. The widget identifies itself as an assistant, and the header states that answers come from your content. Under the EU AI Act, systems that interact directly with people must make that clear at first interaction — we build it in rather than leaving it to you to remember.

Sprosimo is in development and is not selling anything yet, so there is no trading entity to name at this point. It will be published on the Legal Notice page before any payment can be taken — you will never be asked to pay a company you cannot identify.

A fair question to ask any small vendor. Your content and every conversation are exportable as JSON at any time from your account page — no proprietary lock-in. If we wound down we would give at least 60 days' notice and keep exports available throughout.