CelerSearch – AI-Powered Instant Search with Meilisearch

Yes. Two options:

• Self-host on your VPS: free, but you handle updates, backups, and SSL. Upstream guide.
• Managed at celersearch.com: from $12/mo, provisioned in seconds, updates and backups handled.

Three, all running this plugin with AI Semantic Search enabled: an Outdoor Shop (WooCommerce), an Art Gallery (custom post type), and a Cinema. Overview at celersearch.com/demos.

Meaning-aware search. CelerSearch stores a vector embedding for each document and blends semantic similarity with keyword relevance on every query (hybrid search). Searching for keep my coffee hot on the summit surfaces a vacuum flask even though its copy never contains those words. It is optional and enabled per index; with it off, keyword search works exactly as before.

Only for AI Semantic Search. Regular keyword search never calls any AI service. You enter the key per index (Index Settings, AI Search); it is stored server-side, shown masked, used only to generate embeddings, and never exposed to visitors.

OpenAI’s text-embedding-3-small (the default: 1536 dimensions, fastest and cheapest) and text-embedding-3-large (3072 dimensions, most accurate). Developers can register other embedding sources via the celersearch_embedder_sources filter.

You pay OpenAI directly, per token, on your own API key. CelerSearch makes one embedding call per document at index time (and again when a document changes), plus one per search query. At the default model’s $0.02 per 1M tokens (OpenAI pricing as of mid-2026; text-embedding-3-large is $0.13) that stays small: indexing 10,000 typical products costs roughly $0.05 to $0.15 one time, and 100,000 searches cost a few cents. Each document’s embedded text is capped by the Template max bytes setting (default 3000 bytes, roughly 750 tokens), so costs are bounded. Set the semantic ratio to 0 and no embedding calls are made at all.

No. It works with any Meilisearch 1.13 or newer (or 1.3+ with the vectorStore experimental flag), self-hosted included. The CelerSearch Cloud AI plans are simply sized for vector search workloads.

Yes, once. Embeddings are generated at index time, so after enabling AI Search run a rebuild; from then on, new and edited content is embedded automatically as it syncs. Every index-settings tab shows an impact badge telling you whether a change applies instantly or needs a rebuild.

An optional assistant that answers visitor questions in natural language, using your indexed content. For each question it retrieves the most relevant documents from the index, has an AI model compose a short answer from them, and streams the reply live with numbered sources. Place it as a floating widget on the pages you choose, inline with the [celersearch_chatbot] shortcode, or via the Chatbot block.

Semantic search improves the ranked result lists your visitors already use (search page, autocomplete, shop filters). The chatbot answers questions conversationally instead of returning a list. They are independent features: use either, both, or neither.

No. It retrieves supporting content with regular keyword search by default; enabling semantic search on the same index helps it find the right documents for vague or conversational questions.

It is instructed to answer only from the documents retrieved for each question, to never invent products, prices, or facts, and to say it does not know when your content has no answer. That grounding keeps it reliable in practice; still, test it with the built-in console before going live.

OpenAI (gpt-4o-mini by default), OpenRouter, or any OpenAI-compatible endpoint, self-hosted ones included. Alternatively, a Meilisearch server with the chat completions feature enabled can run the conversation natively. As with semantic search, you bring your own API key and it stays server-side.

You pay your AI provider per token, on your own key; CelerSearch adds no fees. With the default gpt-4o-mini, a typical answer costs a small fraction of a cent, and built-in per-visitor rate limits (5 questions per minute and 100 per day by default, configurable) keep usage bounded.

Each question sends the visitor’s message and short excerpts of the retrieved content to the AI provider you configured, nothing more. By default CelerSearch stores nothing. Turn on question logging to keep anonymous question statistics, or “Preserve conversations” to store full transcripts (attributed to the logged-in account when signed in) with bounded retention; the settings include suggested privacy-policy wording you can adopt.

Yes. Dedicated Products and Orders index types, plus AJAX-faceted shop & archive filters.

Yes. Create a Users index and member search becomes instant: the wp-admin Users screen, an optional front-end member directory built with Views, and the BuddyPress members directory (public profile fields are indexed and searchable, respecting each field’s visibility setting). Privacy first: only a strict allowlist of fields is indexed, and email addresses stay out unless you explicitly opt in.

No. Search queries go to Meilisearch, which responds in under 50ms.

With fallback enabled (default), search reverts to native WordPress search automatically.

Yes. Open CelerSearch → Indices, edit the index, and toggle Content Splitting (chunk size is configurable per index).

For site-wide overrides you can also set these in wp-config.php:

define( 'CELERSEARCH_SPLIT_POSTS', false );
define( 'CELERSEARCH_CONTENT_MAX_SIZE', 3000 );

Rebuild the affected indices after changing either.

1. CelerSearch → Views: create a View, pick an index, configure behavior.
2. Create a WordPress page and embed the View via the CelerSearch block or [celersearch] shortcode.
3. CelerSearch → Settings → Search Areas: on “WordPress Public Search”, enable “Replace Search Results Page” and pick the page.

They solve different problems and often work together.

• Search Areas (Settings → Search Areas) intercept WordPress search where it already happens (the frontend search, admin list search, REST API search, WooCommerce shop filters) and route it through Meilisearch. No template changes; you pick which index each context uses.
• Views (CelerSearch → Views) are a UI primitive for building custom search experiences you embed via the CelerSearch block or the [celersearch] shortcode. A View bundles the input, autocomplete, filters, results layout, and sorting as one configurable unit you drop onto any page or widget area.

Typical setup uses both: create a View-powered search page, then enable “Replace Search Results Page” on the WordPress Public Search area and point it at that page. The Search Area handles the interception, the View renders the page.