M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation

The associated BGE reranker models serve as cross-encoders that jointly encode query-document pairs, providing more accurate relevance judgments than bi-encoder similarity at the cost of increased computation.

“The associated BGE reranker models serve as cross-encoders that jointly encode query-document pairs, providing more accurate relevance judgments than bi-encoder similarity at the cost of increased computation.”

The BGE-2 reranker jointly encodes the query and each candidate document, providing more accurate relevance scores than bi-encoder similarity alone.

“The BGE-2 reranker [Chen et al., 2024] jointly encodes the query and each candidate document, providing more accurate relevance scores than bi-encoder similarity alone.”

Cognis's BGE-2 reranker provides substantial improvements on multi-hop reasoning (+11.1% F1 over the Zero Entropy alternative) despite adding only 20-50ms latency.

“Cognis's BGE-2 reranker [Chen et al., 2024] provides substantial improvements on multi-hop reasoning (+11.1% F1 over the Zero Entropy alternative) despite adding only 20-50ms latency.”

The combination of temporal intent detection, time-window scoring, and BGE-2 cross-encoder reranking creates a powerful pipeline for time-sensitive queries.

“The combination of temporal intent detection, time-window scoring, and BGE-2 cross-encoder reranking [Chen et al., 2024] creates a powerful pipeline for time-sensitive queries.”

Cognis has a hybrid retrieval pipeline using RRF fusion (70% vector + 30% BM25), explicit temporal boosting for time-sensitive queries, and a BGE-2 cross-encoder reranker for final result refinement.

“A hybrid retrieval pipeline using RRF fusion [Cormack et al., 2009] (70% vector + 30% BM25), explicit temporal boosting for time-sensitive queries, and a BGE-2 [Chen et al., 2024] cross-encoder reranker for final result refinement”

Recent embedding models such as BGE-M3 natively support both dense and sparse representations, underscoring the importance of hybrid retrieval.

“Recent embedding models such as BGE-M3 [Chen et al., 2024] natively support both dense and sparse representations, underscoring the importance of hybrid retrieval.”

The study uses the BAAI/bge-base-en-v1.5 encoder as the retriever to index both queries and memory items for efficient retrieval and fair comparison across methods.

“For efficient retrieval and fair comparison across methods, we utilize the BAAI/bge-base-en-v1.5 (Chen et al., 2024) encoder as the retriever to index both queries and memory items.”