RAG Reranker Showdown: Small Cross-Encoder Pushed Correct Chunks Out of Top 5
Affectionate_Slip654 · reddit · 2026-10-26
The author benchmarked four rerankers (Cross-Encoder, LLM-based, Cohere, Jina) on 1,507 chunks from payment-industry PDFs in a full RAG pipeline. Key findings: similarity ≠ relevance; a local mini cross-encoder rewarded word matching and pushed both correct answer chunks out of Top 5; an LLM listwise reranker understood context and ranked them 1st-2nd; Cohere lifted an answer from rank 15 to 1. Cohere and Jina offered the best accuracy/speed balance, at the cost of per-call pricing and sending docs to third parties. Code repo and video walkthrough are public.
More from coding & agent
- A successful tool call doesn't mean the agent made the right decision — djgoel123 · 2026-10-26
- OpenClaw ships v2026.10.7 with 61 PRs from 54 contributors — heyneighbor · 2026-10-26
- TokenSpeed's Three-Language Stack: C++ for Correctness, Rust for Agents, Python for Glue — zhyncs42 · 2026-10-26
- IBM's open-source MCP gateway ContextForge unifies MCP, A2A and REST APIs behind one endpoint — tristanbob · 2026-10-26
- Orchestrating Sonnet subagents can cost more than Opus 5.5 on medium effort — Odd-Shift-640 · 2026-10-26
- 10 open-source GitHub repos that give your AI agent a chat UI, by stars — victor_explore · 2026-10-26