01
Problem
Educational questions span exact terminology, semantic concepts and multi-step learning goals, making a single retrieval strategy brittle across formats and query types.
02
Role
Designed and implemented the retrieval pipeline, query transformations, service interface and offline evaluation workflow.
03
Architecture
- 01
BM25 plus BGE-M3 dense and sparse retrieval - 02
Weighted fusion and BGE reranking - 03
Parent–child chunks with PDF, PPTX, DOCX and OCR parsing - 04
Direct, HyDE, subquery and step-back routing
04
Key contributions
- Combined lexical and semantic evidence through weighted hybrid retrieval and reranking.
- Built route-specific query transformations for different question structures.
- Integrated document ingestion, mock interviews and personalized learning plans behind streaming APIs.
05
Evaluation
- A 30-question RAGAS study recorded 97.1% context recall, 95.0% context relevance, 88.7% answer relevance and 83.9% faithfulness.
06
Technology
07