The web exists to be talked to, not just read

The vision: the web should talk, not just be read

For two decades the web has worked the same way: a person lands on a page, scans it with their eyes, navigates a menu, types a term into a search box, and clicks until they reach the right information. It is a model where the human does the work, not the machine. Yet the information already sits inside the site; what is missing is a layer that can give it back in human language, the moment it is asked, by voice.

That is exactly where RAQS starts: a website should be able to talk with its visitor. When a visitor says 'how many sizes does this come in', 'when will it ship', 'how do I return it', or 'show me the best plan for me', they should get a direct, correct answer out loud — without reading the whole page. That is the shortest answer to what RAQS is: the voice of your site.

The commercial logic is just as clear. Content that cannot be found is a customer that is lost. A visitor who gets an instant answer stays longer, trusts more, and converts more. So we designed RAQS not as a 'toy widget' but as the conversion and support layer of a site.

The technology we built

Behind RAQS is an architecture built for production, not just for a quick demo. We built the voice pipeline on Pipecat: a visitor's audio is carried over LiveKit, a low-latency realtime layer, transcribed by Azure Speech, answered by an Azure OpenAI gpt-4o-mini brain, and spoken back again through Azure Speech. Barge-in (talking over the assistant) and natural pauses are standard.

What the assistant 'knows' is strictly limited to your content — RAQS does not make things up. We crawl your site (sitemap + pages; a headless browser for JS-rendered SPA sites; clean text extraction with trafilatura; robots.txt respected), or you paste text/FAQ or upload PDF/Word/TXT/MD/CSV. Content is chunked, embedded with Azure embeddings (text-embedding-3-small) and stored in pgvector.

At query time a hybrid retrieval runs: semantic vector search and Postgres full-text keyword search are fused with RRF (Reciprocal Rank Fusion), and a similarity threshold drops irrelevant results. If the answer is not in your content, RAQS says it does not know. On e-commerce, schema.org JSON-LD on product pages (name, brand, price, stock, category) is extracted too, so 'is X in stock, how much?' is answered correctly.

Not just answers, but action

RAQS is more than a question-and-answer box: it 'sees' the live page and can act on it. It guides the visitor to the product or page they want, shows the relevant item when they say 'show me this', and triggers actions like add-to-cart. To give your brand a face, you can also pick a 3D mascot that roams the page instead of the classic orb — the flagship deer, the universal 'Spark', stylized characters and more than 30 animals. The mascot treats the page's real DOM elements as terrain, running, jumping and leading the visitor to a target.

We balanced that action power with safety. Sensitive operations like payment, account changes and deletion always require the user's explicit confirmation; a smart safety layer makes that boundary mandatory. The assistant never takes an irreversible action on your behalf without consent.

Why voice, why now

Voice is the most natural human interface; keyboard, mouse and touchscreen were always patches over that naturalness. For years voice interfaces stayed weak for two reasons: recognition was not good enough, and the 'brain' behind it could not hold context. Both have changed. Today's speech recognition and language models make it possible to talk to a real site, in real time, at low latency.

The timing is not an accident for us. Large language models can now reliably call tools and stay in context; realtime audio transport layers (like LiveKit) have matured in the browser; cloud speech services have reached production quality. When those three waves break at once, a talking site stops being a research project and becomes a product you can install in one line of code. RAQS was born at that intersection.

The team and our background

The team behind RAQS spent years building products in web and education technology. We worked with content, users and multi-tenant systems at scale, and learned in the field what it really means for one platform to serve hundreds of different customers at once without their data mixing. That is why we designed RAQS as multi-tenant with row-level security (RLS) isolation from day one; each customer's site, knowledge base and conversations belong only to them.

The same background shaped our product decisions. We made our crawler SSRF-guarded, blocked abuse with an origin allowlist and Cloudflare Turnstile, kept infrastructure in Azure's EU region, and built with KVKK and GDPR in mind. The backend is end-to-end Python/FastAPI, the voice pipeline runs on Pipecat, and the marketing site is static Next.js — all live behind Cloudflare under a single domain, raqs.ai.

Our business model reflects that same clarity: a managed/hosted cloud, pay-as-you-go usage plus a base plan, and a free trial. We carry the operational weight, not you; you add a single script line and RAQS handles the rest.

<script async src="https://raqs.ai/v1/raqs.js"
  data-raqs="YOUR_SITE_KEY"></script>