Yemen Opportunity Navigator
A bilingual retrieval-augmented generation system for discovering opportunities using semantic retrieval, BM25, hybrid ranking, reranking, citations, and systematic evaluation.
100%
Recall@5
0.9526
RAGAS Score
5,802
Knowledge Chunks
01
The Problem
Opportunities relevant to Yemeni students and professionals are distributed across many sources, making discovery slow and requiring users to search repeatedly across different websites.
02
The Solution
A multilingual RAG application that indexes structured opportunity content and combines dense semantic retrieval with lexical retrieval, reciprocal-rank fusion, reranking, and source-backed answer generation.
03
My Role
Sole developer. I built the ingestion and retrieval pipeline, multilingual embeddings, hybrid search, reranking, RAG evaluation, authentication features, bilingual interface, and deployment.
04
Architecture
Opportunity sources
Document processing and chunking
5,802 indexed knowledge chunks
E5 multilingual embeddings
Chroma vector retrieval
BM25 lexical retrieval
Hybrid RRF fusion
Cohere reranking
Top retrieved evidence
Answer generation with citations
05
Key Highlights
06
Results
5,802 knowledge chunks indexed
30-question evaluation golden set
Recall@5 = 100%
RAGAS average = 0.9526
Context precision = 1.000
Context recall = 1.000
Faithfulness = 0.9875
07
Challenges
Supporting both Arabic and English retrieval consistently
Balancing semantic and keyword-based retrieval
Ensuring answers remain grounded in retrieved evidence
Measuring retrieval quality separately from answer quality
08
What I Learned
Hybrid retrieval can outperform relying on dense search alone.
Evaluation must test retrieval and generation independently.
High-quality citations significantly improve trust in RAG systems.
Technology