RankGPT Dify Plugin

LLM-based RankGPT reranker plugin for Dify, supporting OpenAI, Google Gemini, Anthropic, DeepSeek, and Grok.

Author: ki3nd
Type: model
Github Repo: https://github.com/ki3nd/rankgpt-dify-plugin
Github Issues: issues

Overview

This plugin adds a rerank model provider to Dify and uses an LLM to reorder retrieved documents by query relevance.

Features

• Supports OpenAI, Google Gemini, Anthropic, DeepSeek, and Grok providers
• RankGPT-style permutation reranking
• Sliding-window reranking for longer document lists
• Rank-based pseudo-score output (1/(rank+1))

Configure In Dify

When configuring the rankgpt provider in Dify:

Field	Description
provider	openai, gemini, or anthropic
model	e.g. gpt-4o-mini, gemini-2.0-flash, or claude-sonnet-4-20250514
openai_api_key	OpenAI API key (OpenAI only)
openai_base_url	Optional, for OpenAI-compatible endpoints (OpenAI only)
gemini_api_key	Google AI Studio API key (Gemini only)
anthropic_api_key	Anthropic API key (Anthropic only)
deepseek_api_key	DeepSeek API key (DeepSeek only)
deepseek_base_url	Optional, for custom DeepSeek endpoint (DeepSeek only)
grok_api_key	xAI Grok API key (Grok only)
grok_base_url	Optional, e.g. https://us-east-1.api.x.ai/v1 (Grok only)
window_size	Sliding window size (default 20)
step_size	Sliding step size (default 10)
max_doc_words	Max words per passage before truncation (default 300)

How Reranking Works

1. The plugin builds a RankGPT-style prompt with the query and indexed passages.
2. The LLM returns a ranking order (e.g. [3] > [1] > [2]).
3. The plugin parses the order, removes duplicates and invalid indices, and applies a safe fallback for missing ones.
4. Results are returned to Dify as RerankResult.

Notes

• LLM-based reranking — latency and cost depend on model choice and document count.
• For large document sets, tune window_size and step_size to avoid oversized prompts.
• score_threshold applies to a rank-based pseudo-score, not a true relevance probability.