In plain words
Web pages often put facts in tables, lists, and infoboxes, where headings explain what the entries mean. Search systems can find the right page but give a language model chunks that miss those facts or their context.
Tessera is a layer that revisits retrieved pages and packs relevant, labeled passages into space the reader already has. In tests on three text and screenshot systems, it improved pooled answer accuracy, though the evaluation covered English Wikipedia.
The research
When a search system finds a useful webpage, will its reader receive the answer evidence with the labels and headings needed to understand it? The authors traced evidence through a retrieval pipeline and found that some evidence stored in its chunks never reached the reader.
Tessera rebuilds retrieved pages into labeled evidence units, then packs relevant units into an existing reader slot without changing the context budget.
In a source-to-reader trace, quoted evidence reached the reader for 59.5% of questions whose cited pages supported their reference answers. For 16.5%, it was stored but missing from the reader’s input.
When researchers supplied a complete stored evidence chunk on 715 affected questions, accuracy rose from 49.1% to 78.5%. In the main comparison, a model reader answered from either the top three retrieved items or the top two plus Tessera’s evidence block in the same third-slot budget.
A grader model judged answers against references. Pooled accuracy increased across three tested text and screenshot pipelines, supporting the idea that delivering page evidence with its context can help.
The main evaluations use English Wikipedia, with pages rich in tables and infoboxes. They do not establish how Tessera handles the open Web’s layout tables, client-side rendering, or noisier markup.
What the evidence shows
- Finding a page is not enough
Web retrieval-augmented generation gives a language model material from searched pages to support its answers. But webpages put facts in tables, infoboxes, lists, and captions as well as prose, and labels or headings often explain what those facts mean.
Fixed chunks or screenshot tiles can separate the evidence from that context, or leave structured evidence out before it reaches the reader.
- Delivering evidence with context
Tessera is a training-free evidence-delivery layer. It treats a ranked list as a way to locate pages, reopens their HTML, and builds small units such as table rows, list items, and paragraphs with the page title, section path, or table headers needed to understand them.
It selects units relevant to the question and uses one existing reader slot for them, leaving the index, retriever, reader, and context budget unchanged.
- Tracing the evidence gap
The authors traced quoted evidence through a hybrid text pipeline to locate where it was lost. Among 5,883 questions whose cited pages supported their reference answers, evidence reached the reader for 59.5%; for 16.5%, it was in stored chunks but missing from the reader’s input.
On 715 questions in this latter situation where one stored chunk contained all the evidence, replacing the third slot with that chunk raised accuracy from 49.1% to 78.5%. This oracle intervention shows the potential value of delivering the evidence, not Tessera’s own result.
- Comparing matched reader budgets
In the main comparison, the reader received either the pipeline’s top three items or its top two plus a Tessera evidence block in the third item’s budget. At the same context budget, pooled answer accuracy rose by 5.0 percentage points on the authors’ hybrid text pipeline, 6.2 on PixelRAG’s text pipeline, and 5.6 on PixelRAG’s screenshot pipeline.
A separate ablation found a larger gain for units with titles, section paths, and headers than for bare units, supporting the role of added structure in that tested setup.
- Extending to repeated search
The layer also improved a multi-step search agent’s answer accuracy on two tests requiring information across sources. The gain was larger on FRAMES than on MuSiQue, whose questions mostly draw on single paragraphs of lead-section prose.
Together, the results support evidence delivery across the tested text, screenshot, and repeated-search setups. The evaluation uses English Wikipedia, so it does not establish how well the method handles different web layouts or rendering.
Deep
- Original paper authors
- Peichun Hua, He Sun, Zirui Zhuang, Qi Qi, Yiming Qiu
- Submitted to arXiv
- License
- arXiv non-exclusive