Prompt Card
goal: Replaced the encoder-based reranker with a decoder that generates the reranking output.
target model: microsoft/phi-4 [Q4-K_M: 9.05GB]
refine model: GPT‑4o (ChatGPT)
available input: 8k tokens
description
This prompt defines a decoder-based ranking evaluator designed to reorder candidate documents from a RAG retrieval pipeline based on semantic relevance. It mimics cross-encoder scoring behavior while ensuring strict output format compliance.
core tasks
- Analyze query intent (definition, method, comparison, factual detail).
- Rank docs by semantic coverage, not keyword overlap.
- Enforce redundancy filtering and penalize near-duplicates.
- Produce cross-encoder-like scores (0.0–1.0) with realistic gaps.
scoring policy
- 0.95–1.00 → Perfect intent match.
- 0.75–0.89 → Strong relevance.
- 0.40–0.70 → Partial relevance.
- <0.30 → Low/no relevance.
- Maintain steep score gaps; avoid uniform decrements.
output format
json
{
"rank": [doc_id, ...],
"score": [float, ...],
"ignore": [doc_id, ...]
}
Top comments (0)