Talk Intermediate MIT First Talk

GRAG Research Presentation

Review Pending

Session Description

I propose Graded Retrieval-Augmented Generation (GRAG) to address the “document zoo problem,” where conventional RAG systems retrieve plausible passages from the wrong documents. My approach preserves document boundaries by first routing a query to the most relevant document, then retrieving chunks within it and expanding to neighboring chunks when needed to recover surrounding evidence. I also introduce an ingestion-time image analyzer that converts visual content into textual descriptions while preserving their original locations, avoiding query-time multimodal processing. I will evaluate GRAG against flat-index RAG baselines and new architectures on heterogeneous documents, focusing on faithfulness, answer relevance, and context precision where it outperform.

Key Takeaways

Contribute my research and knowledge of IR (Information Retrieval).

References

https://lnkd.in/p/ghb2AX3a

Session Categories

Introducing a FOSS project or a new version of a popular project

Talk License: MIT

Speakers

Tushar Soni SDE Intern | Futurelab Studios

I’m a b.tech student specializing in machine learning, passionate about advancing the frontiers of language and speech intelligence.

My expertise lies in language models, from mathematical foundations to low-level implementation. I’ve built strong skills in model optimization, efficient conversion workflows, and practical deployment pipelines, and my focus increasingly centers on the post-training of language models, where I work on refining, adapting, and tailoring LLMs for targeted applications. I’m also diving into inference compilers and high-performance serving systems to further elevate model efficiency, throughput, and scalability.

I've three years of experience in visual designing, working both as a freelancer at Vividly (2022/25) and as an employee at Cimpress (Feb 25 - Aug 25). My journey began back in 9th grade, driven by a strong passion for creativity and design - Behance.

https://iam-tsr.github.io/

Reviews

No reviews yet.