The Reasoning Bottleneck in Graph-RAG: Structured Prompting and Context Compression for Multi-Hop QA
Yasaman Zarrinkia, Venkatesh Srinivasan, Alex Thomo · 2026
Read The Reasoning Bottleneck in Graph-RAG: Structured Prompting and Context Compression for Multi-Hop QA by Yasaman Zarrinkia, Venkatesh Srinivasan, Alex Thomo (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-14045.txt.
Description
Zarrinkia, Srinivasan, and Thomo show that weak reasoning, not weak retrieval, is the main bottleneck in Graph-RAG question answering. In tests of the KET-RAG system, the gold answer sits in the retrieved context for most questions, yet final accuracy lags far behind. Most errors trace to reasoning slips. Two fixes help: SPARQL-style chain-of-thought prompts and graph-walk compression of the context. With both, a small open Llama model matches a much larger baseline at about one twelfth of the cost.
Catalog record
| CREATOR | Yasaman Zarrinkia, Venkatesh Srinivasan, Alex Thomo |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38150 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.14045