A support chatbot that answers from your documents — and admits when it can't.
Customer-facing RAG on your help center, product docs and policies. Every answer is retrieved before it is written, cited back to the source, and refused outright when the documentation doesn't cover the question.
Deflection and languages from a live iGaming deployment (AI chatbots); retrieval and hallucination figures from the Enterprise RAG Challenge.
Support chatbots fail in public.
An internal search tool that gives a bad answer wastes ten minutes. A customer-facing bot that gives a bad answer creates a commitment you didn't make, in writing, to someone who screenshots it. That difference is what drives the whole design.
Decision-tree bots answer the wrong question
Scripted flows only handle what someone anticipated. Anything phrased unexpectedly falls through to "let me connect you to an agent" — which is the outcome the bot was bought to prevent.
An ungrounded LLM invents policy
Point a model at your brand voice without retrieval and it will confidently state refund windows, warranty terms and prices that don't exist. Confident and wrong is the worst possible failure in front of a customer.
The documentation changes; the bot doesn't
Terms get updated, a product ships, a promotion ends. If the index isn't refreshed automatically, the bot keeps answering from last quarter's policy long after support has moved on.
Escalation loses the conversation
Handing off without context forces the customer to explain everything a second time. That single moment is where most of the goodwill the bot earned gets spent.
Retrieve first. Answer second. Refuse when needed.
The bot never generates from memory. Each turn goes through the same path, and the interesting part is what happens when retrieval comes back weak.
Your content becomes the knowledge base
Help center articles, product documentation, policy pages and resolved support tickets are parsed with structure intact and indexed. Past tickets matter more than people expect — they contain the phrasing customers actually use, which is rarely the phrasing the documentation uses.
Hybrid retrieval and reranking on every turn
Dense embeddings handle paraphrase and typos; keyword retrieval catches SKUs, order references and exact product names. Candidates are merged and reranked by a cross-encoder, so what reaches the model is a short list of genuinely relevant passages rather than a large loosely-related dump.
Answer in the customer's language, from the same index
Multilingual embeddings mean one knowledge base serves every language you support — the live iGaming deployment answers in 7 languages without maintaining seven separate content sets. Documentation written in one language can answer questions asked in another.
Grounded generation with a linked source
The reply is written only from the retrieved passages and carries a link to the article it came from. Customers who want to verify can; customers who don't get a shorter answer than the article would have given them.
Refusal and escalation as designed behaviour
When retrieval confidence is low the bot says the documentation doesn't cover it and hands over — carrying the full conversation so the agent doesn't restart from zero. This is the same schema-constrained refusal that produced 0 hallucinations in the Enterprise RAG Challenge, where the system declined on the 18% of questions the source documents couldn't answer.
Continuous evaluation against a golden set
Real customer questions with known-correct answers, run as a suite on every change. Escalated conversations feed back in as new eval cases, so the gaps customers find become the tests that stop them recurring.
A deployed bot, on the channels your customers already use.
Web widget, WhatsApp, Telegram, Slack — the retrieval layer is the same, the surface is wherever your traffic is.
Indexed knowledge base
Your help center, docs, policies and resolved tickets ingested and indexed, with automatic re-indexing so published changes reach the bot without a rebuild.
Multi-channel deployment
One backend, multiple front doors. Web widget plus messaging channels, with conversation state shared so a customer can switch surfaces without losing the thread.
Escalation with full context
Handoff into your existing helpdesk carrying the transcript, the retrieved sources and why the bot escalated — so the agent opens a conversation, not a ticket stub.
Deflection analytics
What was asked, what was answered, what escalated and why. The escalation reasons are the useful part: they tell you which documentation to write next.
Evaluation suite and shadow period
A golden question set plus a run on real traffic with human review of every response before the bot goes live to customers.
Guardrails you control
Topics the bot must never answer, phrasing it must never use, and a hard boundary at anything commitment-shaped — refunds, legal terms, pricing exceptions — that routes to a human by rule.
Pricing lives on the service page
Multi-channel RAG chatbots run $8,000–$18,000. The full breakdown — single-channel, multi-channel and enterprise tiers, plus ongoing API costs and the delivery process — is on the AI chatbots service page rather than restated here.
The retrieval engine behind it, measured three ways.
A public benchmark for accuracy and refusal, and two production deployments for what happens when a real corpus meets real users.
1st place, >90% top-5 hit rate, 0 hallucinations
The Enterprise RAG Challenge, on annual reports, at roughly $0.08 per answer. Schema-constrained output meant the system refused on the 18% of questions with no answer in the source instead of guessing. Not a client engagement — a public, comparable result.
Read the case →Multi-tenant RAG, DIN / EN / DWA norms preserved
Two production tenants — a concrete-products manufacturer and a regional municipal water utility. Hybrid retrieval with a cross-encoder reranker, norm references preserved per chunk, and 23+ golden eval cases per tenant.
Read the case →~150× faster document search for ~500 users
The internal counterpart: a terabyte-scale archive made answerable behind SSO with fine-grained permissions. Same retrieval architecture, pointed inward instead of at customers.
Read the case →The questions that come up first.
What stops it from promising something we can't deliver?
Three layers. It only writes from retrieved passages, so it can't invent a refund window that isn't documented. Low retrieval confidence triggers refusal rather than a guess. And commitment-shaped topics — refunds, legal terms, pricing exceptions — are routed to a human by rule regardless of what retrieval returns.
How much traffic will it actually deflect?
The published reference is 80%+ on a live multilingual iGaming deployment across 7 languages, with typical deployments landing in the 75–85% range. The honest caveat: deflection is a function of how well your documentation covers what customers ask. A thin help center caps the ceiling, and the analytics will show you exactly where.
Do we need to rewrite our help center first?
No. Retrieval works on what exists, and resolved support tickets fill a lot of the gaps because they contain real customer phrasing. The escalation analytics then tell you which articles are worth writing — which is a better ordering than guessing before launch.
Which channels are supported?
Web widget, WhatsApp, Telegram and Slack are the ones already deployed. Anything with an API can be added — the retrieval backend is channel-agnostic, so a new surface is an integration rather than a rebuild.
How does handoff to our agents work?
The bot escalates into your existing helpdesk with the full transcript, the sources it retrieved and the reason it escalated. The customer doesn't repeat themselves, and the agent can see immediately whether the bot failed to find the answer or the answer wasn't there.
What does it cost?
Multi-channel RAG chatbots run $8,000–$18,000, with single-channel and enterprise tiers either side of that. Ongoing API costs are typically $0.01–$0.05 per conversation. The AI chatbots service page has the full breakdown and the delivery process.
Bring your ten most-asked questions.
A 30-minute call. We check them against your existing documentation and tell you honestly how many a grounded bot could answer today, and which ones need something written first. No slides.
Book a call →Response within 24h · Bucharest, Romania