Imported from faionfaion/faion-network (skills/faion/knowledge/ml-engineering/reranking-models/AGENTS.md). Install upstream with npx skills add faionfaion/faion-network --skill reranking-models. Copyright stays with the author.

Reranking Models

Summary

One-sentence: Picks a reranker (cross-encoder vs Cohere/MixedBread API) for two-stage retrieval based on latency, cost, and self-host constraints.

One-paragraph: Two-stage retrieval: fast ANN retrieves top-50 candidates, then a reranker (cross-encoder or API service) rescores them to return top-5 for generation. Choice between local cross-encoders (ms-marco-MiniLM, bge-reranker) and API services (Cohere rerank-3, MixedBread) is driven by latency budget, cost, accuracy targets, and self-host requirements.

Ефективно для: інженерів, які обирають конкретну реранкер-модель для two-stage RAG: cross-encoder локально vs Cohere/MixedBread API.

Applies If (ALL must hold)

• Two-stage RAG retrieval already designed; need to pick the actual reranker model.
• Have constraints on cost-per-query, latency p95, or self-hosting.
• Need to benchmark candidate rerankers on a labeled set.

Skip If (ANY kills it)

• Single-stage retrieval is enough (recall target met without rerank).
• All-in on a managed retrieval product that includes reranking — pick rules don't matter.

Prerequisites

Assumes Loaded

Content (load on demand)

Task Routing

Templates

Scripts

Related

• [[reranking-two-stage]]
• [[reranking-pipeline-integration]]
• [[reranking-diversity-mmr]]

Decision tree

The mandatory tree at content/06-decision-tree.xml picks reranker type by traffic, self-host policy, and recall need. Each leaf references a rule id from 01-core-rules.xml.