PPF Research

Developed Precision Propagation Filtering (PPF), a method for improving retrieval precision in Retrieval-Augmented Generation (RAG) architectures, integrating a linear regression relevance estimator and neural inference layer for dynamic information expansion.

The Problem

Most RAG pipelines suffer from retrieval errors between 15–20%. These occur when too much irrelevant context overwhelms the model, or too little context leaves it guessing. GraphRAG's entity-relationship approach helps in some cases but can underperform when the knowledge graph doesn't align with query intent.

How PPF Works

PPF adds an adaptive filtering layer between the retriever and LLM:

Predictive Precision Model - A lightweight model predicts how much precision the retrieval step should apply. Some queries need breadth (multiple chunks), others need laser precision (1–2 chunks).

Dynamic Context Control - Adjusts retrieval dynamically to prevent "context flooding" (too many tokens) or "context starvation" (not enough). Uses a logarithmic cap function to keep context growth controlled even in massive vector DBs.

Reranking Layer - Integrates with rerankers like Cohere Rerank to ensure only the highest-quality chunks reach the LLM.

Technical Details

• Target hallucination rate: <2%
• Optimal performance with 256-token chunks
• Model-agnostic and vector DB agnostic (Neo4j, Pinecone, Weaviate, Qdrant, FAISS)
• Plug-and-play with existing ingestion pipelines

Results

Validated performance on BEIR benchmark datasets. Tested on Neo4j GraphRAG, with plans to expand to standard RAG with cohesive hierarchical segmentation for broader applicability.

If you want to check it out, the code is on GitHub.