Semantic and Arabic-aware search
Matching intent rather than exact words, and handling Arabic, transliteration and misspellings as customers actually type them.
The problem
Most catalogues are too big to browse and searched with a box that only matches words. A customer who types a description rather than a product name finds nothing, the recommendations are the same best-sellers for everyone, and the merchandising team has no way to influence any of it without a developer. The result is a store where the products that sell are the ones on the first page.
Who this is for
- E-commerce and merchandising managersControl over what surfaces, without raising a ticket
- Retail and marketplace operatorsMore of the catalogue actually reachable by customers
- Marketing teamsSegments built from behaviour rather than from assumption
What people use it for
Personalised recommendation
Related, complementary and next-purchase suggestions from behaviour, with rules the merchandising team controls.
Segmentation and lifecycle
RFM and behavioural segments driving campaigns, reactivation and retention rather than one message for everyone.
Merchandising control
Boosting, pinning and suppressing products by campaign, margin or stock — without a code change.
What it needs to work
- Your product catalogue, with whatever attributes exist — enrichment is part of the work
- Behavioural events: views, searches, cart actions, purchases and returns
- Order history, long enough to contain repeat behaviour
- Your storefront platform and how it can be extended
How it works
-
Understand the catalogue
Products embedded by description, attributes and imagery, so similarity is by meaning rather than by shared keywords.
-
Understand the customer
Behaviour built into a profile that updates within the session, because intent today is not intent last month.
-
Search by intent
Query understanding that handles Arabic, transliteration and misspelling, with results ranked by relevance and availability.
-
Recommend with rules on top
The model proposes, the merchandising rules dispose — stock, margin and campaign constraints are applied last.
-
Measure against holdout
A control group that sees no personalisation, so the uplift is measured rather than assumed.
How we deliver it
-
Catalogue and data review
Attribute quality decides how good search and recommendation can be. Enrichment is usually the first real task.
-
Search first
It is the highest-intent surface in any store and the fastest to show a measurable difference.
-
Recommendation and segments
Added once search is stable, with the holdout in place before anything is claimed.
-
Merchandising controls
Handing the levers to the team who owns the outcome, which is what stops this becoming a black box.
Where it runs
- Integrated with your existing storefront rather than replacing it
- Private cloud, or on-premises where catalogue and customer data must stay in country
- API-first, so search and recommendation can be used in an app as well as on the web
Security and governance
- Personalisation from behaviour you are entitled to use, with consent respected
- Merchandising rules always able to override the model
- Uplift measured against a holdout rather than claimed
- Customer data retained per your policy, not per what the model would prefer
Timeline
Search over a well-attributed catalogue is quick and usually the first measurable win. Recommendation needs enough behavioural history to be better than a best-seller list — with a small catalogue or thin traffic, popularity is genuinely hard to beat, and we will say so rather than ship a model that underperforms a sort order.
What you receive
- Semantic search integrated into your storefront, Arabic and English
- A recommendation service with merchandising rules on top
- Behavioural segments usable by your marketing tools
- Holdout measurement showing the uplift, or showing there is none
- A merchandising console your team operates without us
Related work
Published projects where we did this.
What this does not do
Recommendation needs behaviour to learn from. Below a certain catalogue size or traffic level, a well-ordered best-seller list is genuinely competitive, and we would rather tell you that than sell a model. Personalisation also cannot fix a catalogue with poor attributes or images — enrichment comes first, and it is unglamorous work that determines everything downstream.
Questions we are asked
Do we have to replace our platform?
No. Search and recommendation are integrated as services into what you already run. Replacing a storefront to gain better search is a very expensive way to get better search.
How will we know it worked?
A holdout group that never sees personalisation, running from the start. Without it, any uplift claim is unfalsifiable — including ours.
About the figures on this page
This page describes capability and method. It does not publish accuracy figures, throughput numbers or delivery dates, because those depend on your data, your systems and your scope — and a number published here would be wrong for most readers. You get them, in writing and against your own data, at scoping.
A first call is a technical conversation, not a pitch: what you have, what you need, and whether this is the right approach at all.
Key capabilities
- Retail
- E-commerce
Built on the Unified Intelligence Layer
Every InsAI product runs on the same four-stage backbone.
- 1
Data Integration
ERP · IoT · BIM · CRM
- 2
AI Models & Predictive Engines
Forecasting, detection, optimization
- 3
Automation & AI Agents
Acting on predictions, end to end
- 4
Real-time Dashboards & Decision Systems
From the floor to the boardroom
Smart E-Commerce App
Personalized shopping + recommendations.