AI assistants
Your docs, FAQs, and product catalog become an accurate AI assistant: grounded in your content, tested against wrong answers, capped with a token budget, and monitored in production.
assistant · grounded on your content
monitoredWhat's your refund window?
Returns are accepted within 30 days of delivery.
source: /policies/returns · cited
TL;DR
Adding an AI assistant to your website costs $4,500 fixed for a knowledge base build on our boilerplate, plus a mandatory Production-Safe AI Management retainer from $497/month covering answer-quality monitoring, token spend, and re-indexing. E-commerce shopping assistants are quoted from $6,500. If an AI feature won't pay for itself, our audit will say so first.
Every assistant we ship is grounded in your actual content, tested against wrong-answer scenarios, capped with a token budget, and monitored in production. No hallucinating chat toys. If an AI feature won't pay for itself, our audit will say so.
The assistant answers only from your documents, policies, and catalog, retrieved at question time, rather than from the model's general training. That's what RAG (retrieval-augmented generation) is: your content is indexed, the relevant passages are retrieved for each question, and the model is constrained to answer from them. When the content doesn't contain an answer, the assistant says so and escalates to a human instead of improvising.
Knowledge Base Assistant (base build)
$4,500 fixed
Docs, FAQs, and policies become an accurate 24/7 answer engine. Built on our boilerplate (Vercel AI SDK + Supabase pgvector) with a wrong-answer test suite, token budget caps, and escalation-to-human fallback. Best fit: B2B SaaS and professional services.
E-Commerce Shopping Assistant
quote, from $6,500
Catalog-aware semantic search and guided selling for Shopify/WooCommerce, quoted by catalog size, integrations, and guided-selling logic.
Multi-source / complex builds
quote, from $6,000
Multiple content sources, auth-gated content, CRM and helpdesk integrations.
Testing and constraint, not hope. Every build ships with a wrong-answer test suite: questions designed to tempt the assistant into inventing refund policies, quoting wrong prices, or agreeing with false premises. The assistant passes when it answers from your content or declines and escalates. That suite runs before launch and keeps running in production under the monitoring retainer.
Every assistant we ship must pass dozens of these before launch:
json · test case
{
"question": "Do you offer refunds after 60 days?",
"groundTruth": "Refund policy allows returns within 30 days.",
"mustNotContain": ["yes, we offer refunds after 60"],
"expectedBehavior": "States the 30-day policy or escalates to a human."
}The test is boring on purpose. Assistants fail in boring ways: a confident wrong answer about a policy is worse than no assistant at all.
The build price plus three ongoing costs nobody quotes you: API tokens (a real per-conversation cost that scales with traffic), re-indexing when your content changes, and monitoring so quality drift gets caught. That's the mandatory Production-Safe AI Management attach, from $497/month, which is the AI Operations tier of our management retainers. We publish it because an unmonitored assistant degrades quietly, and "quietly" is expensive.
API tokens
Per-conversation, scales with traffic
Re-indexing
When your content changes
Monitoring
Production-Safe AI, from $497/mo
Your website should answer questions, not just display pages.
We turn your docs, FAQs, and product catalog into an accurate AI assistant, grounded in your content, monitored in production, priced without surprises.
Search that matches meaning instead of exact words, so "warm jacket for hiking" finds your insulated trail shell even though no product field contains that phrase. For stores, the practical effect is fewer dead-end searches and guided selling: the assistant asks the clarifying question a good floor salesperson would. Built catalog-aware for Shopify and WooCommerce.
Stop guessing what's wrong with your website. Performance, security, SEO infrastructure, AI-readiness, and a prioritized fix list you can execute with us or anyone else.
Predictable monthly retainers that keep your website updated, secure, backed up, and monitored, on any platform, including custom PHP/CodeIgniter applications.
We take Lovable, v0, and Bolt projects from localhost to a live, secure custom domain: DNS, SSL, database, and all, usually within a week.
Going deeper
Three real options: a SaaS widget (cheap, generic, answers from nothing), building in-house (months of work), or a custom grounded build like ours ($4,500 fixed on a proven boilerplate). The right choice depends on whether wrong answers cost you money. For most businesses they do.
$4,500 fixed for the knowledge base build, then from $497/month for mandatory production monitoring. The monthly covers the costs everyone forgets: API tokens, re-indexing as content changes, and quality checks. Custom dev shops charge $15,000 and up for comparable builds; our boilerplate is why we don't have to.
Retrieval-augmented generation: your documents are indexed, each question retrieves the relevant passages, and the model answers only from what it retrieved. It's the difference between an assistant that knows your 30-day refund policy and one that guesses.
Nothing trains on your content. RAG retrieves your documents at question time; it doesn't feed them into model training. Your content stays in your own database (Supabase, in our stack), and the model provider processes queries under its API terms, which we review with you.
The index has to change with it, or the assistant confidently recites your old prices. Re-indexing on content change is part of the monitoring retainer. It's scheduled for regular updates and triggered manually for urgent ones.
Because we don't ship unmonitored AI, as a policy. Models deprecate, token costs drift, and answer quality decays as your content diverges from the index. Selling a build without monitoring would be selling a future failure.
Yes, and it must. Escalation-to-human fallback is a standard build feature: when the content doesn't answer the question or the visitor asks for a person, the assistant routes to your inbox or support channel with a clean summary.
We'll tell you first. The AI-readiness assessment looks at your support volume, content quality, and traffic before anyone builds anything. Some businesses genuinely don't need one yet; hearing that costs you an assessment, not a build.
