RAG Evaluations

This section collects all research and evaluation work related to Retrieval-Augmented Generation (RAG) at Unique. RAG is the core retrieval paradigm underlying our AI products, and improving it is a continuous area of investment.

The work here spans foundational benchmarking of our standard retrieval setup, explorations of graph-based retrieval approaches (GraphRAG, LightRAG), and empirical comparisons of model and system performance within RAG pipelines. Each page represents a concrete research effort — with methodology, findings, and strategic conclusions.

What is RAG?

Retrieval-Augmented Generation is a technique that enhances large language model (LLM) responses by first retrieving relevant documents from a knowledge base, then providing those documents as context for answer generation. This grounds the model in real, up-to-date data rather than relying solely on its training knowledge.

At Unique, our RAG pipeline combines vector search and full-text search (combined/hybrid search) with optional reranking to surface the most relevant chunks from customer document sets before passing them to the LLM.

Research areas

• Retrieval quality assessment — Measuring recall and ranking metrics (NDCG) across retrieval strategies.

• Graph-based retrieval — Exploring knowledge graph augmentation via GraphRAG and LightRAG to handle complex aggregation queries.

• Agentic retrieval — How agentic search (UniqueAI-Chat) compares to and extends standard RAG for multi-step reasoning tasks.