Benchmarks

This experimental variant (⁠Echo-DSRN-v0.1.3-Embed-Exp⁠) reaches an average 0.753 Spearman correlation across 7 STS tasks, competitive with dedicated bidirectional encoders (like MiniLM variants) despite its causal recurrent heritage.

Resources Footprint

Because of the O(1) memory core and bounded 128-token local attention, it runs at ~220 sentences/sec on pure CPU with zero GPU dependency and near-instant load times.

Tasks & Active Testing

• Multilingual RAG (⁠pgvector⁠): Intensive validation in PostgreSQL over English, Italian, and German corpora, leveraging the dense 2048-dim slow-state vectors (⁠mean_c_all⁠).
• Semantic Search: Asymmetric retrieval and cross-lingual passage matching