Thesis Proposal — Formulated Idea

Status: draft v1 for supervisor discussion Author: Mohammed Alozaibi University of Aden — Faculty of Computer and Information Technology Date: 2026-08-22

1. Working Title

Primary (EN):

Enhancing Arabic E-Commerce Product Search through LLM-Based Query Expansion over a Hybrid MySQL–Elasticsearch Retrieval Architecture

Alternative titles (pick per supervisor preference):

1. Bridging the Lexical and Semantic Gap in Arabic Product Search: A Hybrid Retrieval Framework Combining BM25, Dense Vectors, and LLM Query Expansion
2. A Comparative Study of LLM Strategies for Arabic E-Commerce Query Understanding in Hybrid Search Systems
3. Shorter / punchier: Smart Arabic Product Search: LLM-Powered Query Expansion on a Hybrid Search Engine

Arabic title (if your university requires it):

تحسين البحث عن المنتجات في المتاجر الإلكترونية العربية باستخدام توسيع الاستعلامات بنماذج اللغة الكبيرة ضمن بنية بحث هجينة تجمع بين MySQL و Elasticsearch

2. One-Paragraph Elevator Pitch

Arabic shoppers type short queries that rarely match how products are described online. A search for ثلاجة fails when the catalog says براد (dialect synonym) or ثلاجه (orthographic variant), and searching أرنب misses products listed as ارنب. This thesis designs, implements, and rigorously evaluates an e-commerce search pipeline that combines (a) a hybrid retrieval architecture — MySQL as the transactional source of truth, Elasticsearch as a fast searchable projection supporting both lexical BM25 and dense vector search with Arabic-aware normalization — and (b) LLM-based query expansion, where one or more large language models generate synonyms, dialectal variants, and rewrites of the user's query before retrieval. The work measures how much each layer improves retrieval quality (NDCG/MRR/Recall) and what latency/cost it adds, comparing several open and commercial LLMs.

3. Problem Statement

E-commerce search is the primary product-discovery tool; its failure directly costs revenue. For Arabic this failure rate is much higher than for English because:

Classical keyword engines handle some of #1–#3 via normalization/stemming but cannot handle #4–#6 semantics; dense embeddings handle paraphrase but are weak on exact SKUs, rare brands, and codes. Neither solves Arabic alone. Meanwhile, industrial evidence (Taobao, Amazon, Walmart — see literature review) shows LLM query rewriting/expansion closes exactly this gap for other languages — nobody has published a rigorous study for Arabic product search.

The gap this thesis fills

1. No public benchmark for Arabic e-commerce product search relevance.
2. No systematic evaluation of LLM query-expansion strategies (synonym generation vs. rewrite vs. hypothetical documents) in Arabic.
3. No reference architecture integrating relational data (MySQL), lexical+dense hybrid retrieval (Elasticsearch), and LLM expansion under realistic latency constraints for Arabic.

4. Research Questions

• RQ1 (Effectiveness): To what extent does LLM-based query expansion improve retrieval effectiveness (NDCG@10, MRR@10, Recall@100) for Arabic e-commerce queries compared to (a) plain BM25, (b) BM25 + rule-based Arabic normalization/synonyms?
• RQ2 (Strategy comparison): Which LLM expansion strategy works best for Arabic product search: (i) synonym/dialect term generation, (ii) full query rewrite, (iii) HyDE-style hypothetical document generation? How do results vary across LLMs (open Arabic-centric models like ALLaM/Jais/AceGPT/Qwen vs. commercial GPT/Claude/Gemini)?
• RQ3 (Hybrid architecture): Does combining lexical (BM25) and dense vector retrieval with rank fusion (RRF) outperform each individually for Arabic product search, consistent with findings for English?
• RQ4 (Cost/latency): What latency and monetary cost does the LLM expansion step add, and can caching + small local models make it production-viable?
• RQ5 (optional, dialect robustness): How do gains differ between MSA queries and dialectal queries?

RQ5 is what turns this from an engineering project into research — keep it if time allows.

5. Hypotheses

• H1: LLM query expansion significantly improves NDCG@10 over normalized-BM25-only baseline (expected +5–15%, based on English-language results in Jagerman et al. 2023 and LLM-QE 2025).
• H2: Hybrid BM25+kNN+RRF beats either single retriever (consistent with Cormack 2009, Bruch et al. 2022).
• H3: Arabic-centric open LLMs produce better dialect expansions than general multilingual models of similar size.
• H4: With response caching, median added latency of LLM expansion stays < 150 ms for head queries.

6. Proposed Contributions

1. An Arabic e-commerce search benchmark: either translate/adapt the ESCI dataset into Arabic, extend the emerging arabic-ecom-search-bench, or build a new annotated dataset (queries + products + relevance grades) — publishable artifact in itself.
2. An end-to-end reference implementation (open source): MySQL → Elasticsearch sync, Arabic analyzer configuration, hybrid retriever, pluggable LLM-expansion service.
3. The first systematic empirical comparison of LLMs for Arabic query expansion across strategies, models, and query types.
4. Practical guidelines: which combination to deploy given a latency/cost budget.

7. Scope & Delimitations

In scope

• Text (query → product title/description/category/brand) retrieval and ranking.
• MSA + major dialect vocabulary (at minimum Egyptian/Gulf/Levantine synonyms relevant to products).
• Offline evaluation with graded relevance judgments; optional small user study.

Out of scope (mention explicitly to survive the defense)

• Personalization / collaborative signals (click-through ranking).
• Image/multimodal product search.
• Training a new LLM from scratch (we fine-tune/prompt existing ones at most).
• Full production hardening (auth, multi-tenancy, horizontal scaling tests beyond a load test).

8. Risks & Mitigations

9. Suggested Timeline (12 months)

10. Why This Will Pass (Novelty Argument)

• Applied novelty: first rigorous Arabic-language evaluation of LLM query expansion for product search (English equivalents exist: Taobao BEQUE 2024, Google QE-prompting 2023 — none cover Arabic).
• Artifact novelty: Arabic product-search benchmark (the field lacks one; ESCI has no Arabic).
• System novelty: documented polyglot architecture (MySQL + ES hybrid lexical/vector + LLM middleware) with measured trade-offs — reusable by regional e-commerce companies.