PseudoSeer: a Search Engine for Pseudocode

PseudoSeer is a novel search engine for academic pseudocode, enabling retrieval over 320,000 algorithm implementations extracted from the arXiv.Using the system's captionreference pairs, we study asymmetric retrieval, matching short queries with a median length of five words against long documents of roughly 300 words composed primarily of natural language with limited LaTeX notation.Our evaluation reveals scaling limitations in embedding models: a 149M parameter encoder outperforms 1.5B parameter alternatives, while BM25 remains competitive with pretrained models.Analyzing attention patterns over 33,000 caption document pairs, we identify two factors driving these results: attention efficiency and attention concentration.Models that significantly attend to sinks or non-discriminative tokens leave less attention for discriminative content, while models with overly diffuse attention fail to form discriminative representations.Guided by these findings, PseudoSeer's embedding model, trained via contrastive learning with efficient attention patterns, outperforms the best pretrained model by 8.7 points.A hybrid approach combining learned embeddings with BM25 reaches 66.5% R@10.PseudoSeer is deployed at pseudoseer.ist.psu.edu as both a practical search system and a benchmark for retrieval evaluation.