My full list of papers are available on Google Scholar.
Key works:
1. Seven Failure Points When Engineering a Retrieval Augmented Generation System
Lays the foundation for building robust LLM applications by presenting 7 failure points when using LLMs for information retrieval (RAG systems). Provides a basis for identifying research problems in this space and presents lessons learned.
Quote from the paper
- “Validation of a RAG system is only feasible during operation, and the robustness of a RAG system evolves rather than designed in at the start.”
2. RAGProbe: An Automated Approach for Evaluating RAG Applications
Builds upon the failure points paper by presenting evaluation scenarios, a systematic and automated method for evaluating RAG systems. The purpose of this work is to provide an automated way to do domain specific evaluation of RAG systems.
Quote from the paper:
- “RAGProbe found more failures (between 45-83% per RAG pipeline, 180 questions) as compared to state-of-the-art.”
Explores the impact a domain has on a RAG system through a case study. This paper highlights that generic evaluation approaches are insufficient compared to human annotations for a narrow domain. However, the annotation schema was designed by humans for humans not machines.
Quote from the paper:
- “We found that LLMs, particularly GPT-4, can precisely answer many single-choice and yes-no questions given relevant context, demonstrating their efficacy in information retrieval tasks. However, their performance diminishes when confronted with multiple-choice and number extraction formats, lowering the overall performance of the model on this task, indicating that these models may not yet be sufficiently reliable for the task.”
4. LLMs for Test Input Generation for Semantic Applications
Presents VaryGen a tool to generate semantically similar data automatically. The key idea was to generate semantically similar text without using similarity measures such as cosine similarity. This was to avoid the problems with embeddings representing a static concept of semantic relatedness. Precursor to RAGProbe.
5. MLGuard: Defend your machine learning model!
The key idea here is to adapt the concept of design by contract to machine learning based systems. Contracts could provide useful but are dependent on knowing what to monitor and how to monitor the ML model. Without that information knowing how to set the contracts is challenging.