RAG 與文件工程
READ Lets Agents Search Structured Long Documents Directly, Reaching 58.8% Accuracy on 51 Questions but Not Significantly Beating BM25
READ replaces fixed chunking and vector Top-K retrieval with lexical search, structure navigation, and section reading, making every evidence-retrieval step replayable. It substantially outperformed dense retrieval on a 780-page financial report, but the sample was small, costs were higher, and the evidence is insufficient to claim that agentic search is superior to traditional lexical indexes.
For table-heavy documents such as financial statements and audit reports, the problem with traditional RAG may not be that the embedding model is insufficiently powerful, but that offline chunking has already destroyed the context around the numbers. A new study examines a [780-page financial report](https://cag.gov.in/uploads/state_accounts_report/account-report-FA-VOL-I-2024-25-069c52aa2b34bf9-63690940.pdf) from Gujarat, India: 86.8% of its content lines are tabular, and its 58,791 numeric tokens contain only 15,960 distinct values. The median distance between monetary-unit headers and value rows is 13 lines, while lakh and crore differ by a factor of 100.
The authors propose [READ](https://arxiv.org/abs/2608.06305), which does not build a vector index. Instead, it exposes three types of deterministic operations through MCP: normalized lexical search, document outline navigation, and reading bounded contiguous sections by line number. An agent can first search for a field name, then expand the reading range upward to include year and unit headers. Every tool input, returned line number, and final citation can be replayed. The design follows the schema-based calling model of the [MCP tools interface](https://modelcontextprotocol.io/specification/2025-06-18/server/tools), but the key lies in the tools’ semantics rather than the protocol itself.
Across 51 manually verified questions, READ achieved 58.8% accuracy, compared with 27.5% when the same agent used a vector Top-K tool. Under matched settings, fixed dense RAG scored 15.7%; increasing retrieval depth raised its best result to 35.3%, still 23.5 percentage points behind. The gap came mainly from single-value lookups, where READ scored 70.8% versus just 12.5% for dense retrieval.
However, the study does not prove that “agentic search wins across the board.” BM25 scored 51.0%, and its gap with READ was not statistically significant on this small sample, while its cost and latency were only about one-third of READ’s. READ’s full operation set also did not significantly outperform the variant without outline navigation. Moreover, PDF conversion can split decimal points and table cells, and no retriever can recover information that has already been lost. Engineering teams should first compare BM25, structure-aware parsing, and auditable section reading before deciding whether they need an expensive agentic loop.