Getting started
How the assistant answers
What happens between a visitor’s question and the streamed answer, and why it stays accurate and affordable.
Routing: the cheapest correct answer wins
Every message goes through a router that tries free and cheap answers before calling the AI model:
- Rules: greetings, thanks, goodbyes, “are you a bot?” in six languages. Instant and never billed.
- Your FAQ: if the question closely matches an FAQ entry (semantic similarity), your answer is returned as written.
- Precomputed answers and semantic cache: common generic questions (“what do you recommend for dry skin?”) are answered from earlier good answers. Cached answers are skipped if any product they mention went out of stock, and they are always rendered with live names, prices and cards.
- AI model: everything else. The model receives the relevant products, your policies and the conversation, and can call tools to search the catalog, fetch product details, read a policy or escalate to a human.
Retrieval: finding the right products
Candidates come from three sources combined: semantic vector search over enriched products, keyword search (names, brands, SKUs) and concern/attribute tags extracted by enrichment. Then come hard filters (price ranges such as “under €30” or “sub 100 lei”) and reranking by stock and relevance.
Guardrails
- Products can only come from your catalog. Product ids in the AI’s output are checked against what it received; anything else is dropped.
- Prices, sale prices and stock in product cards are read live from your catalog, never from the model.
- Health questions get careful, non-diagnostic answers with a recommendation to consult a professional (e.g. pregnancy, allergies).
- Catalog and knowledge content is treated as data, not instructions. Text like “ignore your rules” in a product description has no effect.
- Out-of-scope requests are politely declined.
Languages
The assistant replies in the language the visitor writes in (English, Romanian, Italian, French, German, Spanish), even if your catalog is in one language only. The widget interface follows the page language (see language detection).
Escalation to your team
When the visitor asks for a human, or the assistant cannot help, the widget offers a contact form. The visitor’s message and the full transcript are emailed to your escalation address (Settings), with reply-to set to the visitor. The conversation is marked as escalated in the dashboard.
Usage and billing
AskMerra is priced on the AI work your assistant actually does: answering shoppers (model calls, searches) and learning products (enrichment and indexing of new or changed products). Each piece of work is metered in euros as it happens; your monthly plan fee is included usage, and anything above it is pay-as-you-go at the same rates, paid from prepaid credits.
- Setting up is free: importing the catalog, writing the knowledge base and designing the widget. AI work (answers, enrichment, indexing, the playground) starts once the shop has a paid plan.
- Credits belong to the account that pays for the shop and are shared by all the shops it pays for. Top them up in Usage & billing or on the account page (from €20), optionally with automatic top-up. They never expire.
- Rule answers (greetings, thanks…) are free; FAQ, precomputed and cached answers only cost a tiny search.
- Playground tests use the AI too, so they count as usage, but never as conversations in your analytics.
- When your included usage is used up and pay-as-you-go is off, your credits run out, or your monthly spending cap is reached, questions that need the AI get a polite notice with the contact form, while free answers keep working.
Protection mode