---
title: "Discussion on \"Sent-cosine top-5 kept 100% gold spans at 0.116 token ratio on our 15-query RelayDesk — emb_filter 0.70 retained 13.3%\" | Hashnode"
description: "Discussion on \"Sent-cosine top-5 kept 100% gold spans at 0.116 token ratio on our 15-query RelayDesk — emb_filter 0.70 retained 13.3%\". Author 15 gold answer spans × top-8 RelayDesk packs (seed materials 20260928, API $0). Compared raw concat vs sentence-cosine extractive vs DIY EmbeddingsFilter thresholds vs local Ollama LLM extract."
url: "https://hashnode.com/posts/sent-cosine-top-5-kept-100-gold-spans-at-0-116-token-ratio-on-our-15-query-relaydesk-emb-filter-0-70-retained-13-3/6aba998aa998767264663d9b"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Sent-cosine top-5 kept 100% gold spans at 0.116 token ratio on our 15-query RelayDesk — emb\_filter 0.70 retained 13.3%](https://eworldlive.hashnode.dev/sent-cosine-top-5-kept-100-gold-spans-at-0-116-token-ratio-on-our-15-query-relaydesk-emb-filter-0-70-retained-13-3?utm_source=hashnode&utm_medium=feed)

Author 15 gold answer spans × top-8 RelayDesk packs (seed materials 20260928, API $0). Compared raw concat vs sentence-cosine extractive vs DIY EmbeddingsFilter thresholds vs local Ollama LLM extract.

Sep 28 · 1 min read · eworldlive.hashnode.dev  
Read the article

-   [#contextual-compression](https://hashnode.com/tag/contextual-compression)
-   [#rag](https://hashnode.com/tag/rag)
-   [#embeddings](https://hashnode.com/tag/embeddings)
-   [#sentence-transformers](https://hashnode.com/tag/sentence-transformers)
-   [#prompt-engineering](https://hashnode.com/tag/prompt-engineering)

## Written by

[market](https://hashnode.com/@eworldlive)

@eworldlive

  
[View profile](https://hashnode.com/@eworldlive)

Denser boards for curious AI, tech, and weird-web stories

## Responses

Markdown supported · ⌘↵ to post

No responses yet.