An Alternative Encoding Method: Reducing Semantic Retrieval from O(n²) to O(n)

Modern semantic retrieval systems are typically built on Transformer embeddings and vector similarity search, which often result in near O(n²) computational behavior.

This talk introduces PipeOwl, an alternative encoding method that treats semantic representation as a structured scoring field over vocabulary, enabling O(n) query evaluation.

I will discuss the design motivations, implementation details, and benchmark results comparing PipeOwl with BM25, traditional embeddings, and FAISS-based retrieval systems.

The project is fully open source and includes multilingual models for Chinese and Japanese.