READ: Agentic Document Search Replaces Black-Box Retrieval for Financial Reports
A recent preprint on arXiv (2608.06305) critiques the prevalent retrieval-augmented generation (RAG) framework, which involves chunking, embedding, and top-k nearest neighbor search for lengthy documents. The authors contend that this approach is fundamentally flawed for financial statements, audit reports, and regulatory returns. They present quantifiable data: in a 780-page government financial report, 86.8% of the lines consist of table rows, and numerous nearly identical figures exist within a single embedding space. Additionally, a figure often derives its unit from a header located a median of 13 lines above, potentially leading to significant errors. To remedy these challenges, the authors introduce READ (Reliable Embedding-free Agentic Document-search), which allows for structured document interpretation while circumventing the opaque retrieval model. This paper is categorized as a 'new' announcement on arXiv and underscores the shortcomings of existing RAG systems for structured financial information.
Key facts
- arXiv preprint 2608.06305 proposes READ (Reliable Embedding-free Agentic Document-search).
- The paper critiques the standard RAG design of chunking, embedding, and top-k retrieval for long documents.
- On a 780-page government financial report, 86.8% of content lines are table rows.
- Thousands of near-identical figures compete in one embedding space.
- A figure inherits its unit from a header a median of 13 lines above it.
- Chunk boundaries can separate numbers from their units, causing errors of two orders of magnitude.
- A table-aware chunker fixes the unit problem but leaves 27-30% of numeric chunks without a fiscal-year header.
- READ is an embedding-free agentic approach to document search.
Entities
Institutions
- arXiv