DeepRead: From Fragmented Retrieval to Structure-Aware Agentic Reading

If you work with long PDFs, reports, or academic papers, you probably know the frustration of seeing an AI system “search” a document without really understanding its structure.

So there is a very human question: what if models could navigate a document more like we do?

Limitations of Existing Agentic Search Systems

Figure 1: A Comparison of Search-o1-style Agentic Search and DeepRead on a Toy Case. [Source].

Existing agentic search systems can actively look through long documents, but they still treat everything as flat chunks. They rarely make use of headings, hierarchy or reading order. As a result, they miss information, repeat searches, and suffer from severe context fragmentation.

When people read long documents, the usual approach is to locate the right section first and read the surrounding content contiguously, instead of guessing keywords again and again. Many current agentic search systems still lack this ability.

At the same time, modern OCR and document parsing tools already do a decent job recovering headings, lists, and reading order into structured formats like Markdown. These structural elements are available today and should not be ignored.

This led to an idea: give the model explicit access to document structure so it can reason in a more human way, starting with locating the right part of the document and then reading through it.

DeepRead: Not More Rounds of Search, But Setting the Document Upright First

The idea behind DeepRead is straightforward: restore the document into a form that looks closer to how a person would read it.

DeepRead assumes the raw file is first processed by OCR into a structured Markdown-like format, then each paragraph is labeled with an addressable coordinate (doc_id, sec_id, para_idx). For headings, the system keeps metadata such as children, n_para and n_tok, so the model knows where a passage lives and how large the section is.

Figure 2: This is the DeepRead framework diagram. It takes user questions parsed into Doc Schema as input. Using two tools, LLM can perform multi-turn tool invocations to answer user questions. [Source].

Figure 2 shows this pipeline, from document parsing and index building to multi-step tool calls with an LLM. DeepRead is not a trained agent architecture here.

There are three main pieces that form a synergistic “locate-then-read” loop: TOC(D) for global planning, and a closed-loop pair of tools (Retrieve → ReadSection) for localization and contiguous reading.

• TOC(D) places the document’s headings directly in the system prompt so the model has a clear structural map to work from. The TOC is more than a list of headings: each node also carries children, paragraph counts, and token counts, so the agent can estimate scope and reading cost before deciding how deeply to read.
• Retrieve handles scanning-aware localization. It doesn’t just find the relevant paragraphs; it uses a “scanning window” to return the surrounding context, simulating how a human might skim nearby text, all tagged with precise coordinates.
• ReadSection focuses on continuous, order-preserving reading within a chosen section and a specified span of paragraphs.

Evaluation

Figure 3: Comparison with Different Methods (Accuracy %). Bold indicates the optimal choice, underlined indicates the next best choice. Green text denotes the absolute improvement over the corresponding Search-o1 baseline. [Source].

As shown in Figure 3, DeepRead reaches an overall average of 79.5 percent, which is 10.3 points higher than the 69.2 percent reported for Search-o1. The gains are even more striking on tasks requiring long-range reasoning, such as ContextBench, where DeepRead jumps by a massive 17.0 points.

Figure 4: Cost comparison between correct and incorrect samples. The average number of tool calls and total token consumption are reported across all benchmarks. [Source].

Higher accuracy does not come free. As shown in Figure 4, DeepRead shows higher average tool usage and token consumption than Search-o1 in both the correct and incorrect subsets.

DeepRead gives a straightforward reason for this gap: the baseline appears cheaper largely because of context starvation, meaning it skips portions that should have been read in the first place. Even when the baseline is augmented with passive local expansion, DeepRead’s dynamic, section-level reading still performs better overall.

Thoughts

In my view, the most valuable part of DeepRead not as another small RAG trick, but as a shift in how document structure is treated.

Instead of leaving structure as a side product of indexing, DeepRead turns it into an interface the model can work with directly. With paragraph coordinates, a TOC, Retrieve and ReadSection, the model can actually locate a target area first and then read it in sequence. This feels closer to how people handle real documents than approaches that first construct external hierarchies or graph-like structures, for example RAPTOR, or graph-based methods (like BookRAG and SentGraph).

But I have a concern.

DeepRead is highly sensitive to the quality of document parsing. While it banks on the recent boom of advanced Vision-Language Models (like PaddleOCR-VL) to provide high-fidelity Markdown, the system prompts explicitly warn that parsing errors can turn normal text into headings. Once the topology is reconstructed incorrectly, navigation and section reading will keep moving along a faulty structure, essentially blinding the agent or sending it down a rabbit hole.

Reference: DeepRead: Document Structure-Aware Reasoning to Enhance Agentic Search.

My latest articles: https://aiexpjourney.substack.com/