Skills2026-03-21 · source-backed

Try asymmetric embeddings for RAG.

Issue

2026-03-21

Confidence

source-backed

Sources

Redaction

passed

Story

Use Jina v3's task='retrieval.query' for questions and task='retrieval.passage' for documents — separate LoRA adapters encode through different projection paths, closing the query-document asymmetry gap that single-encoder models miss. Source

LoRA benchmarked against Mistral / Shared entity: LoRA / Same source domain / Shared topic / What happened next / Tension
LoRA patches stay portable across 10 continual-pretraining updates.
Linked by a graph relationship (LoRA benchmarked against Mistral); both cover LoRA; reported by the same outlet (arxiv.org).
DoRA competes with LoRA / Shared entity: LoRA / What happened next
Swap LoRA for DoRA with one flag.
Linked by a graph relationship (DoRA competes with LoRA); both cover LoRA; picks up the LoRA thread on 2026-06-28.
Shared entities / Shared topic / What happened next
Fine-tuning drops to last resort in the 2026 consensus
Both cover LoRA, RAG; overlapping topics (adapter, lora, model, retrieval); picks up the LoRA thread on 2026-06-30.
RAGOCR competes with RAG / Shared entity: RAG / Same source domain / What happened next / Tension
RAGOCR renders retrieved documents as images and beats naive RAG by 15% on one-eighth the input tokens.
Linked by a graph relationship (RAGOCR competes with RAG); both cover RAG; reported by the same outlet (arxiv.org).
LoRA benchmarked against Mistral
Xiaomi Just Open-Sourced a 1-Trillion Parameter Model Under MIT License. It Built a Compiler in 4 Hours.
Linked by a graph relationship (LoRA benchmarked against Mistral).
Unsloth supports LoRA
Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE
Linked by a graph relationship (Unsloth supports LoRA).
LoRA benchmarked against Mistral
Google Ships Gemma 4 Under Apache 2.0. Four Models, Arena #3, and the License That Changes Everything.
Linked by a graph relationship (LoRA benchmarked against Mistral).
Shared entity: RAG / Same source domain / Shared topic / What happened next / Tension
MetaSyn shows agents recover under 53% of relevant literature despite 91% retrieval recall.
Both cover RAG; reported by the same outlet (arxiv.org); overlapping topics (document, retrieval).

Source trail

Source

Entities

LoRA RAG

Provenance

Canonical issue: Ramsay Research Agent — Saturday, March 21, 2026
AI generated: no
Story unit: 2026-03-21-try-asymmetric-embeddings-for-rag
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

Try asymmetric embeddings for RAG.

Story

Related stories

LoRA patches stay portable across 10 continual-pretraining updates.

Swap LoRA for DoRA with one flag.

Fine-tuning drops to last resort in the 2026 consensus

RAGOCR renders retrieved documents as images and beats naive RAG by 15% on one-eighth the input tokens.

Xiaomi Just Open-Sourced a 1-Trillion Parameter Model Under MIT License. It Built a Compiler in 4 Hours.

Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE

Google Ships Gemma 4 Under Apache 2.0. Four Models, Arena #3, and the License That Changes Everything.

MetaSyn shows agents recover under 53% of relevant literature despite 91% retrieval recall.

Source trail

Entities

Provenance

Try asymmetric embeddings for RAG.

Story

Related stories

LoRA patches stay portable across 10 continual-pretraining updates.

Swap LoRA for DoRA with one flag.

Fine-tuning drops to last resort in the 2026 consensus

RAGOCR renders retrieved documents as images and beats naive RAG by 15% on one-eighth the input tokens.

Xiaomi Just Open-Sourced a 1-Trillion Parameter Model Under MIT License. It Built a Compiler in 4 Hours.

Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE

Google Ships Gemma 4 Under Apache 2.0. Four Models, Arena #3, and the License That Changes Everything.

MetaSyn shows agents recover under 53% of relevant literature despite 91% retrieval recall.

Source trail

Entities

Provenance