PaperlayerView exact revision on arXiv

Source-linked AI summary

Tessera: Structure-Completed Evidence Units for Web Retrieval-Augmented Generation

Peichun Hua, He Sun, Zirui Zhuang, Qi Qi, Yiming Qiu, Shinan Liu, Yunming Xiao

arXiv:2610.37080v1cs.IR

TL;DR

Web RAG pipelines use the same chunks or screenshot tiles for indexing, ranking, and reading, while webpage evidence often depends on structure that those units do not preserve. Tessera rebuilds retrieved pages into structure-completed evidence units and packs them within the existing reader budget. It adds 5 to 6 accuracy points on text and screenshot pipelines and also improves multi-step agentic search.

  • Problem

    Web RAG pipelines couple indexing, ranking, and reading, while structured webpage evidence can be lost or separated from the context that gives it meaning.

    Source: §1 · p. 1

  • Method

    Tessera reopens ranked pages, builds fine-grained evidence units completed with page and structural context, selects query-relevant units, and packs them within the unchanged reader budget.

    Source: §1 · p. 2, §3 · p. 4, §3.3 · p. 4

  • Results

    Tessera adds 5 to 6 accuracy points to text and screenshot pipelines and improves a multi-step search agent by up to 5.2 points.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    The reported gains persist across readers, retrievers, chunk sizes, and reader budgets, while the index, retriever, and reader stay unchanged.

    Source: §1 · p. 2, §6 · p. 9

Abstract

from arXiv · show

Web retrieval-augmented generation (RAG) splits a page into fixed-size text chunks or, in recent visual pipelines, screenshot tiles, and the retrieved item serves at once as the unit of indexing, ranking, and reading. We trace answer evidence from source HTML to reader input over the full English Wikipedia and find that this coupling loses evidence in text and screenshot pipelines alike: evidence present in the source page often fails to reach the reader intact, particularly structured content such as tables and infoboxes, which chunking separates from the headers and context that give it meaning. Stronger extraction, context expansion, and larger reader budgets leave the loss in place. We present Tessera, a training-free evidence delivery layer that decouples what a pipeline retrieves from what its reader reads. Given the pages behind any ranked list, Tessera rebuilds their HTML into fine-grained evidence units, each completed with its page title, section path, and table headers, selects the units relevant to the query, and packs them into the reader context at an unchanged token budget. The index, the retriever, and the reader stay as they are, so Tessera applies to text, screenshot, and agentic pipelines. Across three question answering benchmarks, Tessera improves answer accuracy by 5 to 6 points on text and screenshot pipelines, consistently across readers, retrievers, and reader budgets, and outperforms strong HTML-aware and structure-aware RAG baselines under matched budgets. The gains carry over to multi-step agentic search. Our code is publicly available at https://github.com/sarendis56/Tessera.

1 Introduction

Webpages encode facts in structured elements whose meaning depends on surrounding labels and context, but conventional text and screenshot RAG pipelines treat retrieved chunks or tiles as both retrieval and reading units. Tracing those pipelines reveals evidence loss before and after ranking; Tessera instead rebuilds retrieved pages into structure-completed units and delivers them within the existing reader budget.

1 Introduction

  • Conventional web RAG flattens pages and splits them into fixed windows, while screenshot pipelines index rendered pages as tiles.Because the retrieved object also serves as the reading unit, evidence can be separated from the structure that gives it meaning.Source: §1 · pp. 1–2
  • For 7.6% of source-supported questions, evidence never enters the index intact; 81% of these losses are table cells, table headers, or captions.A chunk containing complete evidence can still be missing from the top three; placing it third lifts accuracy on these questions from 49.1% to 78.5%.Source: §1 · p. 2
  • Tessera is a training-free, query-time layer that rebuilds ranked pages into fine-grained evidence units completed with page titles, section paths, and table headers.A cross-encoder selects relevant units, which are packed into one evidence item without changing the reader budget, index, retriever, or reader.Source: §1 · p. 2
  • Across three QA benchmarks, Tessera adds 5 to 6 points to text and screenshot pipelines and outperforms fourteen alternatives under matched budgets.The gain persists across readers, retrievers, chunk sizes, and reader budgets; the paper also reports up to 5.2 points inside a multi-step search agent.Source: §1 · p. 2

2 Where Does Web Evidence Go Missing?

The section examines how text and screenshot pipelines lose evidence and derives design principles for delivering it. Its evidence trace finds losses both after ranking and when structured content fails to survive extraction.

2.1 The Modality Gap Is a Ranking Gap

  • After reranking, text and screenshot pipelines meet: their accuracy differences are −0.4 points on NQ, +1.4 on NQ-Tables, and +2.6 on SimpleQA.Reranking adds 8.9 to 11.8 points to text and 5.2 to 9.5 points to screenshot pipelines; Table 1 reports answer accuracy and gold-article Recall@3.Source: §2.1 · p. 2, Table 1 · p. 3

2.2 Tracing Evidence from Source to Reader

  • Quoted evidence reaches the reader for 59.5% of source-supported questions; for 16.5%, it survives in stored chunks but is lost later in the pipeline.The trace follows evidence through extraction, stored chunks, retrieval candidates, reranking, and the three chunks delivered to the reader.Source: §2.2 · p. 3, Figure 2 · p. 3
  • Replacing the reader’s third slot with a stored chunk containing complete evidence raises accuracy from 49.1% to 78.5% on the affected questions.The intervention covers 715 questions; familiar remedies each add at most 2.2 points, compared with 29.4 points available from the oracle chunk.Source: §2.2 · p. 3, Table 2 · p. 3
  • Extraction loses at least one evidence fragment from every stored chunk for 7.6% of supported questions, with table cells, headers, or captions involved in 81% of those cases.Among quoted table cells found verbatim in chunks, 31% lack the table’s leading column headers; familiar remedies add at most 2.2 points despite the 29.4-point oracle opportunity.Source: §2.2 · p. 3

2.3 Design Principles

  • The design principles are to retrieve pages but deliver finer-grained units, preserve each unit’s structure, parse source HTML at query time, and reallocate rather than expand the reader budget.The top ten reranked chunks reach the gold page for 92.7% of supported questions but deliver evidence for 59.5%; one well-chosen slot is worth 29 points on affected questions.Source: §2.3 · pp. 3–4

3 Tessera

Tessera is a training-free delivery layer that replaces one retrieved item with a same-budget block of structure-completed evidence units. It leaves the index, retriever, and reader unchanged while adapting retrieval results for reading.

3.1 Problem Setting

  • The layer replaces the last reader item with a unit block assembled from pages behind the ranked list, keeping the other top items and the original token-cost ceiling.Tessera reads the query, ranked list, and source-page HTML; the resulting context costs no more than the displaced item.Source: §3.1 · p. 4

3.2 Anchor Pages

  • Tessera collects distinct pages behind the top-ranked items, fetches their HTML, and can cache query-independent units per page.The ranked list is used as a noisy pointer to passages but a reliable pointer to pages.Source: §3.2 · p. 4

3.3 Structure-Completed Units

  • Tessera completes each page unit with its title, section path, and relevant table context so a reader or reranker can interpret it on its own.Block units include paragraphs, list items, definitions, and captions; row units represent table and infobox rows.Source: §3.3 · p. 4
  • A faithful table grid resolves spans, title rows, multi-row headers, group rows, and key–value rows before assigning each row its context and labels.Nested tables inherit their enclosing table’s title and headers, and cell line breaks and list items are preserved.Source: §3.3 · p. 4

3.4 Unit Selection

  • A lexical prefilter narrows about a thousand page units to 100 candidates, then the pipeline’s text reranker scores their completed text against the query.Completed context helps disambiguate which match, season, or entity a row belongs to.Source: §3.4 · pp. 4–5

3.5 Organized Budget-Neutral Packing

  • The packer selects units by score within the displaced item’s budget, then groups them so each context path and header appears once above its rows.The block reads like a table excerpt and fits about 14 lines where repeating context for each unit would fit about 11 units.Source: §3.5 · p. 5

3.6 Plugging into Pipelines

  • Tessera plugs into text, screenshot, and agent search pipelines without changing their prompt, tool schema, or stopping rule.The layer adds no index or offline corpus pass; page parsing takes 15 ms per page on one CPU core, and scoring takes 0.2 GPU-seconds per question on an A100.Source: §3.6 · p. 5

4 Experiments

Across text and screenshot pipelines, Tessera improves accuracy under matched reader budgets, and gains persist against alternative methods, larger budgets, and multi-step search. The analysis links much of the gain to evidence delivery while identifying remaining limits in retrieval and unit selection.

4.1 Setup

  • The evaluation covers 7,851 questions across NQ, NQ-Tables, and SimpleQA on an August 2025 English Wikipedia snapshot, using three pipeline configurations and three readers.The same Tessera configuration was used across experiments and was not tuned per pipeline, reader, or agent.Source: §4.1 · p. 5

4.3 Comparison with Alternatives

  • On the matched-budget hybrid text pipeline, Tessera beats seven retrieval methods by 4.1–6.1 points, RAPTOR by 3.8, contextual retrieval by 4.3, both HtmlRAG variants by at least 5.4, and Docling by 0.8 pooled.Docling gains 4.3 points; Tessera parses a page 26 times faster than Unstructured and over 500 times faster than Docling.Source: §4.3 · pp. 6–7, Table 4 · p. 7

4.2 One Layer, Many Pipelines

  • 5.0, 6.2, and 5.6 points: Tessera raises pooled accuracy on the hybrid text, PixelRAG text, and PixelRAG screenshot pipelines, respectively, under the same context budget.Gains are significant on every dataset and pipeline, and remain between 4.6 and 5.8 points across four robustness checks.Source: §4.2 · p. 6, Table 3 · p. 7

4.4 Accuracy Under a Growing Reader Budget

  • Across 51 pooled comparisons, Tessera beats chunks-only contexts by 1.8 to 6.7 points; its smallest context beats the largest baseline, while mixed allocations outperform units alone.With 1,024-token chunks, 3K tokens with Tessera reach 76.1%, compared with 73.2% for 12K tokens of chunks.Source: §4.4 · p. 7, Figure 6 · p. 8

4.5 Agentic Search

  • On FRAMES, Tessera raises agent accuracy from 59.2% to 64.4%; on MuSiQue, which mostly uses lead-section prose, it gains 1.1 points.With Tessera, the agent averages 2.19 searches rather than 2.41 with chunks and finishes with a shorter context, 11.7K rather than 12.8K tokens.Source: Table 5 · p. 8, §4.5 · p. 8

4.6 Analysis

  • Tessera raises evidence delivery from 72.1% to 81.9% and accuracy from 49.1% to 65.0% on source-supported questions, but the oracle remains 13.4 points ahead.Of 96 answers in the oracle gap, 61 lie outside the anchor set; the remaining 35 are within it, split between the lexical prefilter (20) and selection and packing (15).Source: Table 6 · p. 8, §4.6 · p. 8

5 Related Work

Related work addresses different points in web RAG, from page extraction and visual retrieval to chunking, compression, and agent policies. Tessera instead selects structure-completed units from retrieved pages at query time.

Context compression

  • Unlike post-retrieval refiners that shorten passages, Tessera selects complete structured units and is orthogonal to context compression.It also complements learned extraction, HTML-aware RAG, and standalone-table QA approaches by converting structure into explicit context on small units.Source: §5 · pp. 8–9

Web content and tables in RAG

  • Tessera turns page structure into explicit context on small units, selecting table rows with merged cells resolved from HTML under a budget shared with surrounding prose.It uses no markup tokens or trained models.Source: §5 · p. 9

Retrieval granularity and chunk context

  • Unlike indexing-time granularity methods, Tessera operates at query time on retrieved pages’ HTML and is independent of index granularity.Those methods pay their cost across every page and mostly operate on extracted text that may already have lost structured evidence.Source: §5 · p. 9

Visual document retrieval

  • Unlike screenshot retrieval, which uses fixed-window tiles, Tessera uses top-ranked tiles only to identify pages and builds evidence units from their HTML.Visual retrieval methods include DSE, ColPali, VisRAG, M3DocRAG, and PixelRAG.Source: §5 · p. 9

Agentic RAG

  • Unlike agentic RAG methods that optimize retrieval while accepting tool output formats, Tessera’s self-describing evidence accompanies fewer searches and more correct answers.The cited results describe the agent outcome in §4.5.Source: §5 · p. 9

6 Conclusion

Tessera decouples evidence delivery from retrieval and improves results across text and screenshot pipelines as well as multi-step search. The paper’s implementation materials are slated for release alongside the code.

6 Conclusion

  • Tessera adds 5 to 6 accuracy points across three pipelines and two modalities, holds gains across readers, retrievers, chunk sizes, and budgets, and beats fourteen alternatives at matched cost.For a multi-step search agent, it improves accuracy by up to 5.2 points while issuing fewer searches and reading a shorter context.Source: §6 · p. 9
  • The experiments use English Wikipedia; extending Tessera to open-Web layout tables, client-side rendering, and noisier markup requires a larger, specific corpus.The conclusion identifies this extension as future work.Source: §6 · p. 9

A Implementation and Reproducibility

  • The authors state they will release the code, run manifests, unit builder, and evaluation records; runs used Python with vLLM 0.25.1 on NVIDIA A100 GPUs.The hybrid text pipeline’s implementation details include Trafilatura 2.0.0, dense and BM25 retrieval, and reciprocal-rank fusion.Source: A Implementation and Reproducibility · p. 10

B Evidence Trace Details

The evidence-trace analysis measures how much Tessera closes the oracle gap and contrasts evidence-bearing chunks with arbitrary chunks from the same page.

B Evidence Trace Details

  • Tessera closes 98 of 133 oracle answers (74%) where evidence was stored but not delivered; the oracle leads Tessera by 13.4 points (95% CI 10.2 to 16.7, p=5.4 × 10^-15).Table 7 accounts for the oracle gap; its oracle eligibility uses exact matching while the stage analysis uses tolerant matching.Source: B Evidence Trace Details · p. 10, Table 7 · p. 11
  • The evidence chunk beats an arbitrary same-page chunk by 16.7 to 50.3 points on every dataset, while the arbitrary chunk leaves overall accuracy at baseline.Table 8 reports the same-page control, restricted to questions with a control chunk within 20% of the evidence chunk’s length.Source: B Evidence Trace Details · p. 11, Table 8 · p. 11

C Standard Remedies in Detail

Standard remedies—including improved extraction, expanded context, structure signals, and larger budgets—leave substantial evidence-delivery losses. Learned extraction can recover some lost evidence, but has coverage and processing-cost limitations.

C Standard Remedies in Detail

  • Re-extraction with Trafilatura’s recall-oriented mode doubles coverage of 449 lost-evidence questions but changes accuracy by at most 0.8 points.Recovered text still must win one of three reader slots in a 1,024-token window.Source: C Standard Remedies in Detail · p. 11, Table 9 · p. 11
  • Adding stored neighbors from the ten best pages to the candidate pool and reranking it brings at most 1.5 points.This remedy expands context around hits.Source: C Standard Remedies in Detail · p. 11, Table 9 · p. 11
  • Title prefixes move evidence into the reranker’s top three for 66 of 3,718 questions, and heading paths add three more; table headers push out more evidence chunks than they bring in.Showing title-prefixed chunks to the reader changes accuracy by less than a point in either direction.Source: C Standard Remedies in Detail · p. 11, Table 9 · p. 11
  • Increasing the evidence budget from 3K to 12K tokens raises pooled accuracy from 71.1% to 73.2%, while screenshot context growth from 3 to 14 tiles adds 2.2 points.The larger text budget is four times the reader cost; screenshots still share the text pipeline’s ceiling.Source: C Standard Remedies in Detail · p. 11, Table 9 · p. 11
  • MinerU-HTML covers 275 of 449 lost-evidence questions, but misses 156 table cells—including 97 infobox cells—and would require about 80 A100 GPU-days for the corpus.The authors report that MinerU-HTML classifies infoboxes as boilerplate.Source: C Standard Remedies in Detail · p. 11

D Ablations

The ablations show that completing evidence units with their page and table context is the key source of gain, while faithful table construction and a fixed budget split improve delivery without making results hinge on one serialization.

D Ablations

  • Completed units add 4.6 points and significantly outperform bare units on every dataset; bare units add 1.8 points pooled and nothing on NQ-Tables.Prefixing page titles to chunks adds 0.5 points, while adding the unit block as a fourth item costs about 900 more tokens for only 0.2 points beyond budget-neutral replacement.Source: D Ablations · p. 11
  • The faithful grid corrects labels for a third of table rows and adds 1.4 points on NQ-Tables over DOM-order pairing; grouped packing fits a quarter more lines.The four construction and packing variants stay within 0.34 points pooled, and the grid with table grouping is adopted as source-faithful and most compact.Source: D Ablations · p. 11
  • Jointly ranking chunks and units loses 0.6 points, while replacing all chunks with units loses 1.6 points.A second unit slot, changing anchor-page or prefilter counts, adding sibling rows, and removing duplicated unit text each change accuracy by at most 0.3 points or do not help.Source: D Ablations · p. 11

E Complete Budget Results

The budget sweep reports pooled text-pipeline accuracy across chunking, evidence budgets, and allocation, alongside per-dataset screenshot results.

E Complete Budget Results

  • Table 11 lists every cell of the text budget sweep of §4.4, including 512-token chunks with a 64-token overlap; Table 12 gives the screenshot sweep per dataset.Source: E Complete Budget Results · p. 11
  • Table 12 reports screenshot-pipeline accuracy per dataset when replacing top-k tiles with top-(k−1) tiles and a unit block at the same reader cost.Every gain in the screenshot sweep is significant.Source: E Complete Budget Results · p. 11
  • The text budget sweep includes 512-token chunks with 64-token overlap and varies chunking, evidence budget, and allocation.The table reports pooled accuracy.Source: E Complete Budget Results · p. 11

F HtmlRAG at Its Design Point

At HtmlRAG’s 4K-token design point, pruned HTML alone loses much answer-bearing text, while Tessera leads the best HtmlRAG configuration. The agentic-search details specify a bounded search-tool protocol and a breakdown by reasoning type.

F HtmlRAG at Its Design Point

  • Table 10 covers hybrid-text accuracy ablations, Table 11 pooled accuracy by chunking and allocation, and Table 12 same-cost screenshot tile replacement.Table 10 marks default rows with shading and notes significant differences; Table 11 defines “+1K units” as 1,024 tokens for units and the rest for chunks.Source: Table 10 · p. 12, Table 11 · p. 12, Table 12 · p. 12
  • Tessera leads the best HtmlRAG configuration by 5.8 points (p < 10^-49), and Tessera units alone reach 74.6%.Pruned HTML alone causes the reader to decline 1,305 questions; either HtmlRAG variant after the top-2 chunks scores below chunks alone.Source: F HtmlRAG at Its Design Point · p. 12, Table 13 · p. 12
  • Table 13 compares HtmlRAG and Tessera at a 4,096-token evidence budget and reports reader declines.The evaluation uses all 7,851 questions and the same anchor pages as Tessera.Source: F HtmlRAG at Its Design Point · p. 11, Table 13 · p. 12

G Agentic Search Details

  • The FRAMES agent uses Qwen3.5-4B, at most eight calls to one search tool, and three returned documents per call; Table 14 breaks results down by reasoning type.Each search call runs the hybrid text pipeline on the agent’s query.Source: G Agentic Search Details · p. 12, Table 14 · p. 12

H Robustness of the Main Comparison

Tessera’s gain remains significant across four robustness checks, though sampled residual failures often involve grading or reader errors. Screenshot-context results, processing costs, and SimpleQA examples further characterize its trade-offs.

H Robustness of the Main Comparison

  • Among 50 sampled Tessera failures per dataset, outdated or invalid references or grading errors account for 72% on NQ, 50% on NQ-Tables, and 12% on SimpleQA.Across pipeline failures, reader errors on delivered evidence outnumber missing-evidence errors, 63 to 20 in total.Source: H Robustness of the Main Comparison · p. 13
  • Tessera’s gain is significant (p < 10^-12) across four checks: a 16,384-token output limit, updated-answer grading, supported-reference questions, and 4,851 held-out questions.The held-out set excludes the first 1,000 questions from each dataset used in earlier pipeline experiments.Source: H Robustness of the Main Comparison · pp. 12–13, Table 15 · p. 13

I Unit Blocks in a Screenshot Context

  • At the cost of three screenshot tiles, text and image unit blocks have the same accuracy (p = 0.89), and two of the three tile budgets allocated to units is best.Units without any tile beat three tiles by 4.9 points; with Gemma 4 31B, the text block is 1.9 points more accurate than its image form.Source: I Unit Blocks in a Screenshot Context · p. 13, Table 16 · p. 13

J Processing Cost

  • Tessera never times out, while HtmlRAG and RAPTOR add GPU passes over every page.Table 17 compares the processing each structure-aware method adds, reporting times per page on one CPU core unless noted.Source: J Processing Cost · p. 13, Table 17 · p. 13

K Additional Examples

  • Across 7,851 SimpleQA questions, Tessera gains 620 answers and loses 225; one example recovers an Ig Nobel winner from a year-headed list item.Another example shows the fixed split displacing a third chunk that contained the answer while the unit block did not.Source: K Additional Examples · p. 13, Figure 7 · p. 13

Related papers

Loading 2610.37080v1…