GBrain says vector-only retrieval matches hybrid on top-5 recall at far lower cost
garrytan · x · 2026-07-27
GBrain claims 97.6% recall@5 on the public LongMemEvals benchmark, edging out MemPalace's published 96.6% result on the same dataset.
The screenshot argues that for conversational retrieval, vector-only retrieval is essentially as good as hybrid at top-5, while being much cheaper: around $0.50 per 1,000 queries versus about $3 for a hybrid + LLM expansion setup. The takeaway is that BM25-plus-RRF complexity may not be worth it unless you need higher K or keyword-heavy text.
More from coding & agent
- Anthropic engineer shares a 2-hour workshop on graph engineering for agents — usamawahabkhan · 2026-07-27
- LangChain’s Interrupt workshops focus on deep agents, production and monitoring — LangChain · 2026-07-27
- Cohere launches North Automations for plain-language enterprise workflows — cohere · 2026-07-27
- Google Cloud Model Armor targets prompt injection and data leaks in multi-agent systems — Pavan_Belagatti · 2026-07-27
- Codex gets mocked for serializing eight parallel jobs into an eight-hour loop — chaumian · 2026-07-27
- Cached input makes looped agent workflows cheaper than they look — yunta_tsai · 2026-07-27