Fetching from the wire…
01
02
03
04
05
06
07
08
Skills2026-06-08 · source-backed
Pull 50–100 candidates with a fast bi-encoder, re-score each (query, chunk) pair with a cross-encoder (Cohere Rerank 3.5, Voyage rerank-2.5, BGE v2, Jina v2), pass only the top 3–10 to the model. This fixes the dominant failure where the right chunk is in your top 50 but not your top 5. Practitioners report 20–40% accuracy gains. If your RAG is mediocre, this is the first thing to fix.
Each link below shares sources, entities, or timing with this story.
Jina uses Matryoshka / Shared entities / Same source / Shared topic / What happened next
Linked by a graph relationship (Jina uses Matryoshka); both cover Cohere Rerank, Voyage; cite the same source (Cohere Rerank 3.5, Voyage rerank-2.5, BGE v2, Jina v2).
Shared entities / Shared topic / Earlier coverage
Both cover BGE, Cohere Rerank; overlapping topics (chunk, cross-encoder, each, gain); earlier BGE coverage from 2026-03-19.
Jina uses Matryoshka / Shared entity: RAG / Earlier coverage
Linked by a graph relationship (Jina uses Matryoshka); both cover RAG; earlier RAG coverage from 2026-03-23.
RAGOCR competes with RAG / Shared entity: RAG / What happened next / Tension
Linked by a graph relationship (RAGOCR competes with RAG); both cover RAG; picks up the RAG thread on 2026-08-09.
DeepTutor uses Jina / Shared entity: Jina / What happened next
Linked by a graph relationship (DeepTutor uses Jina); both cover Jina; picks up the Jina thread on 2026-08-14.
Shared entity: RAG / Shared topic / What happened next / Tension
Both cover RAG; overlapping topics (accuracy, query); picks up the RAG thread on 2026-06-19.
Cascade uses BGE / Shared entity: RAG / Earlier coverage
Linked by a graph relationship (Cascade uses BGE); both cover RAG; earlier RAG coverage from 2026-03-14.
Shared entity: RAG / Shared topic / What happened next
Both cover RAG; overlapping topics (accuracy, model, only); picks up the RAG thread on 2026-08-19.