Replacing JoliSearch with Meilisearch in ThirtyBees – early results

[Adik]

Hi everyone! For the last few years I’ve been using JoliSearch module v4.3.28. It’s been a staple in my store, but as my catalog grew to over 10k products, I felt it was time for something faster and more precise, especially for technical search terms. I’m not a hardcore developer, but I’m passionate about making my store run better. 😊

---

Why replace JoliSearch?

JoliSearch has several limitations today:

• It is no longer actively maintained.

• Search relevance is difficult to fine-tune.

• Loose matching often returns hundreds of results.

• The first visible results are not always the most relevant.

• Users may leave the store because they cannot quickly find what they are looking for.

In my case, searching for a popular product type returned over 500-800 products, many only partially matching the intent. That creates noise instead of helping the customer.

For technical stores (industrial hardware, connectors, cables, IPC systems, etc.), this becomes a serious UX and conversion issue.

---

Why Meilisearch?

Meilisearch is a modern, open-source search engine designed specifically for high-performance, real-time search experiences.

https://github.com/meilisearch/meilisearch

Key characteristics:

• Index stored in RAM --> extremely fast response times (often 1–5 ms).

• Built-in typo tolerance and smart ranking.

• Simple and clean REST API.

• Lightweight and easy to self-host (currently running on same VPS as my store).

• Much easier to tune than older search modules.

Native support includes:

• Synonyms

• Custom ranking rules

• Faceted search

• Filtering

• Typo tolerance controls

• Vector search (embeddings support)

---

Automatic typo handling

Meilisearch automatically handles common input problems:

• Minor spelling mistakes

• Missing hyphens (e.g. “usbc” vs “usb-c”)

• Word order variations

At the same time, it allows strict control for technical catalogs:

• Disable typo tolerance for SKU/reference fields

• Limit the number of allowed typos depending on word length

• Keep technical codes exact (e.g. 81271, CA-SASA-12CU)

This is extremely important in stores with many product references and model numbers.

---

Dynamic suggestions & smart autocomplete

The module already includes a live search endpoint and basic fast autocomplete.

Further improvements (some already implemented, others in progress) include:

• Real-time product suggestions with image, price, manufacturer, and reference

• Intelligent grouping (products, categories, manufacturers, feature values)

• Query preprocessing for better intent detection

• Smart result limiting to avoid overwhelming users

Even in its current state, this approach can significantly reduce search exit rates compared to classic result pages.

---

AI integration – OpenAI embeddings & hybrid search

One of the most exciting aspects is semantic search.

Meilisearch supports vector search, which allows:

• Storing product embeddings

• Performing similarity-based queries

• Combining keyword search + semantic similarity (hybrid search)

Using the OpenAI Embeddings API (or local embedding models), we can:

• Generate embeddings from: product name, technical parameters, categories, descriptions

• Store them in Meilisearch

• Enable natural language queries

This enables:

• “Cable for powering laptop via USB-C 100W”

• “Splitter for two devices”

• “Industrial ethernet connector”

The goal is not to replace keyword search, but to enhance it.

---

Current status of my module

After short initial testing, the results are very promising.

Already implemented:

• Custom index (products_pl, products_eng)

• Batch reindexing (500 products per batch)

• Live progress bar in BO

• Live search endpoint

• Synonyms editor (graphical table UI)

• Automatic JSON generation for Meilisearch settings

• Query preprocessing for better intent detection

• Matching strategy control (strict vs fallback)

• Monitoring estimated result counts (to avoid result explosion)

The improvement in relevance compared to JoliSearch is clearly visible, especially in edge cases e.g. “Y-type cables”, where search behavior can now be precisely controlled.

The target is a search interface that behaves more like a modern SaaS-powered discovery engine rather than a traditional e-commerce search box — fast, relevant, visually structured, and intuitive for users (as shown in attached screenshot)

Has anyone experimented with Meilisearch in ThirtyBees yet?

Edited 3 hours ago by Adik

[the.rampage.rado]

Good work!

Sharing is caring! 🙂