AI Chatbots

A CelerSearch chatbot is an AI assistant that answers questions using your own indexed content as its only source. Ask it "which tent fits four people?" and it finds the matching products in your index and writes a short, grounded reply that links them. If your content does not cover the question, it says so instead of inventing an answer.

Create one under CelerSearch → Chatbots, point it at an index, and drop it on your site as a floating widget, a block, or a shortcode.

How it works

A chatbot combines two distinct AI steps. Keeping them straight is the key to understanding what you need to configure.

StepWhat it doesConfigured on
RetrievalSearches your index for the documents most relevant to the questionThe index
GenerationAn LLM writes the answer from those documents, and only thoseThe chatbot

Every question follows the same path:

  1. A visitor asks something.
  2. The chatbot searches your index for the top matching documents.
  3. Those documents are handed to the language model as its only permitted source.
  4. The model writes a concise answer that cites them. If nothing relevant was found, it returns your configured no-answer message rather than guessing.

Because the answer can only draw on what retrieval returned, the chatbot cites its sources and cannot invent products, prices, or facts that are not in your index.

Two different keys

The chatbot's provider key (OpenAI, OpenRouter, ...) powers generation — writing answers. If you also use AI Semantic Search, that embedder key powers retrieval — and it is a separate key for a separate job. A chatbot needs a provider key; the embedder key is optional.

No — a chatbot works without it. Retrieval falls back to keyword search, so as long as your index has content, the chatbot can answer.

But enabling AI Semantic Search on the index makes retrieval far better, and it is the single biggest quality lever for a chatbot:

  • Without embeddings the chatbot finds documents that share literal words with the question.
  • With embeddings it finds documents by meaning. "Something to keep drinks cold on the trail" surfaces an insulated bottle even if that product never says "cold" or "trail."

Visitors ask questions in natural language, so semantic retrieval consistently finds the right content where keyword matching misses. The chatbot's semantic ratio (in its Retrieval settings) tunes the keyword-to-meaning balance, and only takes effect when the index has an embedder.

Recommended setup

Enable AI Semantic Search on the index and rebuild it once so existing documents get embedded, then point your chatbot at that index. New and edited documents embed automatically from then on.

Choosing a provider and model

Chatbots generate answers through any OpenAI-compatible provider:

  • OpenAI — enter your OpenAI API key and a model such as gpt-4o-mini.
  • OpenRouter — one key, hundreds of models across vendors.
  • Custom — any OpenAI-compatible endpoint (a self-hosted model, a gateway) via its base URL.

The model field takes the exact model id your provider expects.

Use the provider's exact model id

Model ids are provider-specific and must match exactly. OpenRouter, for example, uses versioned slugs like anthropic/claude-sonnet-4.6 or openai/gpt-4o-mini — there are no -latest aliases. An id the provider does not recognize makes the chatbot reply with an error. Browse valid ids on your provider's models page.

Setting one up

  1. CelerSearch → Chatbots → Create Chatbot.
  2. General — name it and pick the index it answers from.
  3. AI Provider — choose OpenAI, OpenRouter, or Custom, paste your key, and set the model.
  4. Personality — give it a name, a greeting, optional instructions, and suggested questions.
  5. Retrieval — how many sources to use per answer and the keyword-to-semantic balance.
  6. Placement — enable the floating widget (and where it shows) and copy the shortcode or use the block.
  7. Save, then Test it right in the form before it goes live.

Where it appears

  • Floating widget — a launcher in a corner of the site. Show it everywhere, only on selected pages, or everywhere except some.
  • Shortcode — embed it inline on any page or post with [celersearch_chatbot id="N"].
  • Block — add the CelerSearch Chatbot block in the editor and pick the bot.

Give it an avatar, a launcher label, a custom header title, and an optional proactive "teaser" message under Appearance.

Saved conversations and privacy

By default nothing about a conversation is stored on your server. Two independent, opt-in settings under Limits & Privacy change that:

  • Log questions keeps anonymous statistics (the question and result counts, no identities) so you can spot content gaps.
  • Preserve conversations stores full transcripts, grouped into threads and attributed to the logged-in user when there is one. Read them anytime from the conversations viewer on the chatbot's row in the list.

Preserving conversations stores personal data

Transcripts contain whatever visitors type, which can include personal information, and are attributed to logged-in users. If you enable this, mention it in your site's privacy policy. Threads are pruned automatically and removed if the chatbot is deleted.

Abuse controls (per-visitor question limits, logged-in-only) live in the same section, so a burst of traffic can't run up your provider bill.

Answer engine (advanced)

The Answer engine setting is Automatic by default, which is the right choice for almost everyone. When your search instance supports Meilisearch's native conversational search it is used automatically; otherwise the built-in pipeline handles the exchange. Both ground every answer in your index and stream the reply the same way — the difference is transparent to visitors.

Next steps