Techbanq

AI use case · search

It is not the ranking. It is the morphology

Arabic search fails on the same word written five ways: with and without diacritics, with a hamza someone dropped, in the plural, prefixed. A system that treats those as different strings returns nothing, and users conclude the document is not there.

Serving Saudi Arabia from our Riyadh office and our engineering hubs in Lahore and Karachi

Server racks in a controlled-access facility
  1. 01

    Market

    Saudi Arabia

  2. 02

    Written for

    CIO, Head of Digital

  3. 03

    Decision

    Make internal search find the current document

  4. 04

    Languages

    EN · AR

Enterprise delivery credentials

  • ISO

    ISO/IEC 27001

    ISO/IEC 27001 certified

Certifications & registrations

The fixes

Four changes, in the order that recovers the most

None of these requires replacing the search platform, and the first two are configuration rather than development.

Normalisation

Unifies alef and hamza forms, taa marbuta and haa, strips diacritics and tatweel, and folds Arabic-Indic and Latin digits together.

What the user notices: A query typed the way people actually type it starts returning results.

Stemming and roots

Matches a word to its root and its inflections, so a search for one form finds the document that used another.

What the user notices: Singular finds plural, and the definite article stops being a barrier.

Bilingual synonyms

Maps the English term staff use in conversation to the Arabic term used in the document, and the internal acronym to both.

What the user notices: The policy is found whichever language the searcher was thinking in.

Freshness and ownership

Superseded documents are demoted or removed, and each set has an owner responsible for currency.

What the user notices: The first result is the current version, which is the only result most people read.

Measure first

Twenty queries tell you more than a platform evaluation

Take the twenty things staff most often look for, run them against the current system, and record what comes back and where the right answer ranks. That baseline takes an afternoon and it changes the conversation from "search is bad" to a list of specific failures, most of which have specific causes.

It also settles whether the problem is retrieval at all. A meaningful share of "search does not work" complaints turn out to be permissions — the document is found by the system and hidden from the searcher — or content, where the current version was never uploaded. Neither is fixed by a better engine.

Common causes

Three findings that recur in every estate

  1. 01

    The document exists three times

    A copy on the shared drive, one in the document system and one attached to an old email thread. Search ranks the wrong one first, and the fix is ownership rather than relevance tuning.

  2. 02

    Scanned documents have no text

    A large share of Arabic archives are images. They are invisible to search until they are processed, and no ranking change reaches them.

  3. 03

    Everyone searches for the internal name

    Projects and systems have names used in conversation and never written in documents. A synonym list built from what people type is the cheapest relevance improvement available.

Our role

What we do to a search estate

  1. 01

    Baseline the twenty queries

    Measured before any change, so improvement is demonstrated against a record rather than asserted.

  2. 02

    Fix the analysis chain

    Normalisation, stemming and synonyms configured for Arabic and for your own vocabulary, on the platform you already run.

  3. 03

    Add semantic retrieval where it earns it

    Hybrid search helps on descriptive questions and adds little on document-name lookups; it is applied where the baseline shows it recovers something.

Which document is authoritative is a decision for the function that owns it; search can only surface the answer that decision has already made.

Common questions

What buyers ask

Do we need to replace our search platform?+

Usually not. Most platforms support Arabic analysis properly and are running with defaults nobody configured. Measuring the baseline first tells you whether the ceiling is the platform or the configuration, and in our experience it is the configuration and the content far more often than the engine.

Is semantic search better than keyword search?+

For different questions. Keyword search wins when someone knows the document name or a specific term; semantic search wins when they can only describe what they need. Running both and combining the results is normal practice, and choosing one exclusively is what produces a system that is worse at half the queries it used to answer.

How does this relate to an internal assistant?+

Retrieval is the shared foundation. An assistant that generates answers over a corpus is only as good as what retrieval hands it, so fixing search first improves both — and if search is broken, an assistant built on top of it will produce fluent answers from the wrong documents, which is worse than returning nothing.

Next step

Run it against your own corpus

A measured baseline of what current search returns for the twenty queries staff actually run, and a fix list ordered by what each one recovers.