Asrith Sunke
Research Mentor: Thore Bergman
Mentor Department: Psychology, LSA
Author(s): Alvaro Vega Hidalgo
Session: Session 7 (4:00 PM – 4:50 PM)
Presentation Type: Poster 87
Abstract
Recent work in artificial intelligence has demonstrated the effectiveness of Retrieval-Augmented Generation (RAG) systems in improving the factual accuracy and grounding for language models, particularly in specialized educational domains. Building on these findings, this project investigates how a multimodal, locally deployed RAG pipeline can enhance automated learning tools for bird identification. The research question guiding this work is: How can retrieval systems be optimized to provide accurate, explainable support for beginners learning to identify bird species? To explore this question, we designed a lightweight RAG architecture that integrates Wikipedia-derived ornithological text, Xeno-Canto audio metadata, and structured bird-trait information into a vector database using FAISS and Sentence Transformer embeddings. The system retrieves the most relevant contextual chunks and combines them with a compact Llama-3 model to generate grounded questions. Development involved iterative refinement of chunking strategies, metadata filtering, scoring methods, and prompt design. Preliminary evaluations indicate that section-based chunking and metadata-aware retrieval substantially improve answer precision and reduce hallucination. We expect final results to demonstrate that domain-specific retrieval design can meaningfully enhance model reliability in educational settings. This work contributes an interpretable, scalable approach for building interactive learning tools that combine generative AI with curated scientific knowledge.
