New Features

A/B testing for personalised search

You can now run a controlled experiment to measure the impact of personalisation on your store's search performance. Visitors are split into two groups — a control group that sees standard Typesense ranking and a treatment group that receives personalised ranking based on their click, add-to-cart, and purchase history.

Assignment is deterministic: the same visitor always lands in the same group with no cookie or session storage required. Changing the experiment name in admin immediately reshuffles assignments, letting you start a fresh test without touching the database.

Improvements

Cache stampede prevention (ProfileBuilder)

The 15-minute TTL now has jitter (base 900s + a random 0–120s), so users who start sessions in the same window don't expire in lockstep and trigger the expensive GROUP BY rebuild all at once. On a cache miss, a lock ensures only one request per user rebuilds the profile — concurrent requests for the same key return empty immediately (personalization is a boost, not a hard requirement) rather than dog-piling the database. Late arrivals just pick up the warm cache on the next request.

Async analytics event writes

The bigger pressure point wasn't Redis at all — it was the synchronous MySQL INSERT on every click and search event, which hits connection limits well before the cache layer becomes a concern. That path is now asynchronous: the Track controller publishes the event to a message queue and the request returns as soon as it's enqueued, while a background consumer handles the actual insert. It reuses the existing queue infrastructure, so there are no new dependencies.

Net effect: read pressure is spread across time and deduplicated per user, and write pressure is decoupled from the request lifecycle entirely.

New Features

• Personalization Signals — Search results and category pages now boost products based on each shopper's individual behaviour. Products the shopper has clicked, added to cart, or purchased rank higher for that visitor using Typesense's _eval() scoring formula.
• Personalization works for both logged-in customers (scoped by customer ID) and guests (scoped by session cookie).
• Boost scores are computed server-side at page render and cached in Redis for 15 minutes per visitor — no per-request DB queries.
• Degrades gracefully: visitors with no history see the standard ranking unchanged.

Configuration

New Personalization group under Stores → Configuration → Ayasoftware → Typesense Search:

Database

• Added nullable customerid column (with index) to ayasoftwaresearch_events. Run bin/magento setup:upgrade after installing.
• Analytics tracking now records customer_id for logged-in shoppers, enabling per-customer history that persists across sessions.

Ayasoftware TypesenseSearch — 1.2.1

Bug Fixes

• Fixed search results page not showing any filters
• Fixed price attribute not appearing in the filter sidebar
• Fixed price slider max value being permanently reduced after filtering
• Fixed price slider not resetting when the price filter chip is removed
• Fixed Enter key not applying the typed price range

Improvements

• Price slider now fires search only when the handle is released, not on every pixel drag
• Price handles show a live tooltip with the current value while dragging
• Price range input is now editable — type a range and press Enter to apply
• Improved price slider styling: larger handles, cleaner input, better spacing
• Search results default to 12 items per page with a 12 / 24 / 36 per-page selector
• Filters pre-configured out of the box: Category, Color, Size, Material, Gender, Price

New Features

Let AI Understand Search Intent

• Transform natural language searches into intelligent filters, categories, price ranges, brands, and attributes automatically.

Search Analytics & Recommendations

• Added a Magento Admin search analytics dashboard with store and 7/30/90-day filters.
• Tracks search impressions, product clicks, add-to-cart events, orders, conversion rates, and revenue attribution.
• Added actionable recommendations for zero-result searches, low click-through queries, and frequently searched terms that generate no orders.
• Added daily analytics aggregation for efficient reporting.

AI Search Improvements

• Added AI-powered natural-language filter parsing, including multilingual queries.
• Added provider selection for Claude, OpenAI, and Google Gemini.
• Added provider-specific API key and model configuration shared by query rewriting and AI filters.

Improvements

• Improved search result product rendering and interaction handling.
• Expanded product attribute configuration options.
• Updated the module and Composer package version to 1.2.0.

New Feature: AI-Powered Query Rewriting

This release introduces optional AI-powered search query rewriting using the Claude API. When enabled, search queries on the result page are automatically cleaned and expanded before being sent to Typesense, improving the quality of results for conversational, abbreviated, or misspelled queries.

Examples

• "cheap red running shoes" → removes filler, focuses on product terms
• "tee shirt xl" → expands abbreviations to "t-shirt xl"
• "do you have argus tank" → strips conversational phrasing to "argus tank"
• Fixes common misspellings automatically

How It Works

• Only fires on the search result page (not autocomplete) to avoid latency impact
• Rewritten queries are cached in Redis for 24 hours — repeated queries cost nothing
• If the API is unavailable or times out, search falls back to the original query silently
• When a query is rewritten, the result page shows "Showing results for: [rewritten query]" with an option to search the original term instead

Setup

Go to Stores → Configuration → Ayasoftware → Typesense Search → AI Search, enable query rewriting and enter your Claude API key (available at console.anthropic.com).

Estimated cost: $2–$20/month for most stores depending on search volume and cache hit rate. Uses Claude Haiku (cheapest model, 60 output tokens max).

• Initial release