Master Thesis — Arabic E-Commerce Search with Hybrid Retrieval + LLM Query Expansion

Author: Mohammed Alozaibi University of Aden — Faculty of Computer and Information Technology

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

Documentation map (read in this order)

1. docs/01-thesis-proposal.md — formulated idea: titles, problem statement, research questions, hypotheses, contributions, scope, timeline
2. docs/02-literature-review.md — annotated bibliography by theme with relevance notes and novelty-gap table
3. docs/03-system-design.md — proposed architecture (MySQL → ES hybrid BM25+kNN+RRF, LLM expansion service) + evaluation methodology
4. docs/04-datasets-models-tools.md — datasets, embedding models, LLM candidates, tooling, infra notes

Research library

Downloaded papers (validated PDFs): research/papers/

research/papers/
├── arabic-nlp/            # AraBERT, MARBERT, GATE, ALLaM, Jais, AceGPT, surveys, classics
├── llm-query-expansion/   # Google PRF-prompting, CSQE, LLM-QE, best-practices
├── ecommerce-search/      # ESCI, Amazon Semantic Product Search, Taobao/Walmart industrial systems
└── hybrid-retrieval/      # RRF, weighted-sum fusion, SPLADE

Full source list including link-only/paywalled items: research/SOURCES.md

The one-sentence idea

Arabic product queries fail against keyword search because of orthographic variants (ثلاجة/ثلاجه), dialect synonyms (براد vs ثلاجة), morphology (والثلاجات), and transliterated brands — this thesis adds an LLM query-expansion layer on top of a hybrid BM25 + dense-vector Elasticsearch index (synced from a MySQL source of truth) and measures retrieval quality, latency, and cost across multiple LLMs and expansion strategies.

Status

• Idea formulation v1
• Literature sweep + 24 papers collected
• Discuss RQs/dataset plan with supervisor
• Build benchmark dataset
• Implement baselines (S0–S5 in system design)
• LLM experiment grid (S6+)
• Write thesis