Ricci-Filtration: Boosting Retrieval-Augmented Generation Reranker to Query-Answer Tasks by Discrete Ricci Flow

2026-06-16 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing, Data Science & Analytics · Depth: Expert, extended

Summary

Ricci-Filtration is a novel RAG reranker enhancement procedure that leverages discrete Ricci flow to filter irrelevant document chunks before reranking. By modeling the input query and initial retrieved chunks as a graph, it uses geometric curvature to evaluate chunk importance. The system first filters chunks based on their curvature relative to the query, then a reranker processes the remaining subset. Experiments on datasets like SQuAD2.0 and MultiHop-RAG show Ricci-Filtration outperforms several baseline reranking methods in accuracy, precision, recall, and F1 scores, improving metrics by about 5% in SQuADv2 and significantly boosting accuracy for null queries in MultiHop-RAG (from 60.80% to 83.39%). It uses OpenAI's text-embedding-3-small for embeddings and Llama-3.1-8B-Instruct or gpt-4o-mini for generation, with a default of 20 Ricci flow iterations and a 50th percentile cosine-dissimilarity threshold for graph construction.

Key takeaway

For Machine Learning Engineers optimizing RAG systems, consider integrating Ricci-Filtration as a pre-reranking step. This geometric filtering method can dynamically remove noisy chunks, potentially improving accuracy and F1 scores, especially for single-hop QA tasks. While it introduces computational latency, its LLM-agnostic filtering decision offers a different accuracy-latency trade-off compared to LLM-based rerankers. Evaluate its impact on your specific datasets, particularly for multi-hop reasoning where its current performance may be limited.

Key insights

Discrete Ricci flow can geometrically filter irrelevant RAG chunks, improving reranking performance and reducing noise.

Principles

• Negative Ricci curvature indicates irrelevant chunks relative to a query node.
• Normalized discrete Ricci flow can detect community structures in graphs.
• Geometric filtering can improve answer quality without LLM-based filtering decisions.

Method

Construct a graph from query and chunk embeddings using cosine dissimilarity. Apply finite-time normalized discrete Ricci flow for M=20 iterations. Filter chunks with negative Ricci curvature (high edge weights) relative to the query node. Rerank the remaining chunks.

In practice

• Use precomputed embeddings for graph construction.
• Integrate into existing RAG pipelines for geometric denoising.
• Set Ricci flow iterations (e.g., M=20) for effective separation.

Topics

• Retrieval-Augmented Generation
• Reranking
• Discrete Ricci Flow
• Graph Neural Networks
• Question Answering
• LLM Performance
• Geometric Filtering

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Issue #115 - Reranking in your RAG pipeline — Reranking enhances RAG pipeline precision by reordering initially retrieved documents using a Cross-Encoder.
• WaveFilter: Enhancing the Long-Context Capability of Diffusion LLMs via Wavelet-Guided KV Cache Filtering — WaveFilter uses wavelet transform to identify critical tokens in long sequences, creating a sparse KV Cache to improve Diffusion LLM performance.
• Query-focused and Memory-aware Reranker for Long Context Processing — A new reranking framework uses attention scores for listwise relevance estimation, outperforming existing methods in long context processing.
• Self-Augmenting Retrieval for Diffusion Language Models — SARDI uses DLM tentative tokens as lookahead signals for dynamic retrieval, boosting RAG performance and throughput.
• Disentangling the Strengths of Semantic ID and Item ID Recommendation, Efficient Graph-Based Indexing for Multi-Vector Retrieval, and More! — Generative recommendation models excel at generalization but struggle with memorization, a trade-off addressable via adaptive ensembles.
• RASER: Recoverability-Aware Selective Escalation Router for Multi-Hop Question Answering — RASER reduces multi-hop QA token costs by selectively escalating retrieval based on one-shot RAG features.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.