Ai Agents by White Peaks Solutions SAS
An AI assistant that knows your website — and nothing else.
Ai Agents reads the site you already publish and answers your visitors from it, in the language they asked in. No invented facts, no generic chatbot script: when the answer is not in your content, the agent says so and hands the conversation to your team.
Which plan includes 4K streaming?
The Premium plan includes 4K and two simultaneous streams. Standard is limited to Full HD on one device.
/plansما رقم جوّال المدير التنفيذي؟
هذه المعلومة غير منشورة على الموقع، لذا لا أستطيع الإجابة عنها. يمكنني تحويلك إلى فريق الدعم إن أردت.
Illustration. Every answer cites the page it came from, or the agent refuses.
What it does
One assistant per site, grounded in that site’s own content
Each of your websites gets its own agent, with its own name, tone and knowledge base. The platform is content-agnostic: it knows nothing except what that site publishes, which is exactly what makes its answers trustworthy.
Answers from your content only
Retrieval runs over the pages, PDFs and documents you actually publish, combining semantic and keyword search. A similarity threshold gates every answer before the model writes a word, so an off-topic question is refused rather than improvised — and logged for you to review.
Six languages, natively
Questions are answered in the language they were asked — Arabic, English, French, German, Greek and Spanish — each with its own stemming for keyword search. Retrieval is cross-lingual, so a question in Greek can still be answered from a page you only publish in English. The interface mirrors completely for Arabic, with correct right-to-left layout, and a thread that mixes languages renders each message in its own direction.
Always up to date
Scheduled jobs re-read your sitemap, re-render the pages that changed, index new URLs and drop removed ones. The agent never quotes a page you deleted last week, and you never have to re-upload anything.
Capabilities
Built for real websites, not demo content
Everything below ships as part of the platform. Modules that only some sites need — catalogue recommendations, article summaries, subscription plans, schedules — are switched on per site.
Reads JavaScript-rendered sites
Your Vue or React front end is crawled with a real headless browser, so client-rendered routes are indexed exactly like static pages. Sitemaps are used when you have one, a same-origin crawl of the rendered DOM when you do not.
Incremental, low-cost sync
Each run compares sitemap timestamps, HTTP validators and content hashes, so unchanged pages are never re-processed. Changed pages are re-indexed atomically — retrieval never sees a half-updated document.
A refusal gate, not a guess
Every retrieval is scored before generation. Below the site’s threshold the agent declines and records the question, which turns your unanswered questions into a ranked list of the content you are missing.
Recommendations from your catalogue
Ask for “a drama series” or “something like this title” and the agent suggests items from your own published catalogue — built from the structured data already on your pages — returned as cards with links and artwork.
Summaries of your own articles
Ask what a piece says and the agent loads that article whole — not a handful of search fragments — and summarises it with its byline and publication date. Attribution follows the article: what a source claimed stays a claim, and if only part of a long read fits, the agent says so instead of presenting half a story as the whole one.
Handover to a human
When the agent refuses twice, or the visitor asks for a person, it switches to the handoff mode you configured: a contact form, your existing helpdesk URL, or a webhook that posts the transcript into your own systems.
A distinct identity per site
Name, persona, avatar, welcome message per language and widget colours are set per site and take effect immediately. White-label clients can hide the attribution line entirely.
One line to embed
A single script tag renders the widget inside a Shadow DOM: your stylesheet cannot break it and it cannot leak into your page. It works on desktop, adapts to a full-screen layout on mobile, and runs inside native app WebViews.
Analytics and content gaps
Conversations, answer rate, refusal rate, handoffs, latency, token usage and language split — per site and over time — next to the ranked questions your content does not answer yet.
How it works
From a domain name to a working agent
Setup is a configuration exercise, not an integration project. Nothing about your existing site has to change.
Connect your domain
Add the site in the dashboard and give the agent a name and a tone. We look for your sitemap first and fall back to crawling the rendered DOM, respecting robots.txt and staying on your own origin throughout.
We index what you publish
Each page is rendered, stripped of navigation and boilerplate, split into heading-aware passages and embedded. PDFs, Word documents and manually entered answers join the same index, and structured product or title data becomes your recommendation catalogue.
Embed one line — or share a link
Drop the script tag into your layout, or send visitors to the hosted chat page we publish for the site. Both use the same agent, the same knowledge base and the same language rules.
It keeps itself current
A schedule you choose adds new pages, re-indexes changed ones and removes deleted ones. Every run is recorded with counts and failures, and a failed sync raises an alert rather than silently serving stale answers.
Most sites are fully indexed on the first run; from then on the schedule keeps the agent current without anyone touching it.
Security & compliance
Your content and your visitors’ questions stay on your infrastructure
The platform is designed to run on hardware you control and to hold as little personal data as possible. These are properties of the architecture, not settings you have to remember to switch on.
Self-hosted inference
- The language model and the embedding model run on your own server and are bound to localhost — they are never exposed to the internet.
- No prompt, no page and no conversation is sent to a third-party model provider.
- No per-token bill from an external API, and no vendor that can change its terms under you.
GDPR
- The only personal data stored is conversation text tied to an anonymous visitor identifier — no cookies and no third-party trackers.
- Conversation bodies and unanswered questions are encrypted at rest with AES-256-GCM under a versioned key hierarchy.
- Right to erasure and data export for a visitor are a single action in the dashboard.
- Retention is configurable per company and enforced by a nightly purge.
- A records-of-processing document maps every stored field to its purpose, lawful basis and retention period.
ISO 27001 alignment
- Role-based access control with company- and site-scoped roles, checked on every request.
- An immutable audit log of every administrative action, authentication event and transcript access.
- A documented key hierarchy and rotation procedure; secrets are supplied by the environment, never committed.
- Dependency auditing, cross-tenant access tests and prompt-injection tests run in CI on every change.
- Nightly encrypted backups with a restore procedure that is actually rehearsed.
Use cases
The same pipeline, whatever your site publishes
The agent has no built-in idea of what a “product” or an “episode” is. It learns your vocabulary from your pages, which is why it fits verticals that look nothing alike.
OTT and streaming
Answer questions about plans, devices, regional availability and billing, and recommend titles from your own catalogue — “a drama series like this one” returns published titles with links and artwork rather than a search page.
- Plans and billing
- Device support
- Title recommendations
Support portals
Turn a help centre into a conversation. Visitors get the specific answer from the right article instead of ten search results, and anything the agent cannot answer is escalated to your team with the full transcript attached.
- Help centre answers
- Escalation with context
- Content-gap reports
Developer documentation
Answers grounded in the version of the docs you actually publish, each linking back to the exact page. Configuration and code examples come from your documentation, not from a model’s memory of an older release.
- API reference
- Guides and tutorials
- Source-linked answers
E-commerce
Product questions answered from your product pages, shipping rules and returns policy, with suggestions drawn from the catalogue metadata you already expose — categories, variants and availability as you publish them.
- Product questions
- Shipping and returns
- Catalogue suggestions
News and publishing
Readers ask what a piece says and get a summary of that article and nothing else, with its byline, its publication date and a link back to it. Older coverage stays findable by description rather than exact headline, and a story you never published gets a straight “I cannot find that” instead of a plausible invention.
- Article summaries
- Archive by description
- Byline and date
Pricing & demo
Contact us for pricing
Tell us which site the agent should learn. We will reply with pricing options, an indexing estimate, and a demo agent running on your own content — so you can judge the answers rather than a slide deck.
- We only use your details to answer this enquiry.
- The demo runs on your public content — no credentials or data exports needed.
- No obligation, and no data retained if you decide not to continue.
Prefer email?
Write to us with your domain and we will reply with pricing and the same information.
[email protected]