Ask a typical retrieval system about a number in a financial table and watch what happens. It often answers confidently and wrongly. Not because the model is weak, but because it never really saw the table. It saw a flattened stream of text where rows, columns, and headers had already dissolved into each other.
Text-Only Reading Loses the Document
A document is more than its words. Layout carries meaning. A figure caption belongs to its figure. A total belongs to its column. A heading scopes the paragraphs beneath it. When a document is reduced to plain text, that structure is the first thing to disappear.
- Scanned pages have no extractable text at all, only pixels.
- Tables collapse into runs of numbers with no sense of which value sits under which header.
- Figures, charts, and diagrams carry meaning that text extraction simply discards.
- Multi-column layouts get read straight across, splicing unrelated sentences together.
Watch Out
If meaning is lost when the document is first read, no retriever, reranker, or model downstream can recover it. The damage is already done.
Reading Documents the Way People Do
People do not read documents as a character stream. They see the page. They notice that this block is a table, that region is a figure, this line is a heading. Reading a document visually means giving the system the same starting point: the page as it actually looks, with its structure intact.
Once structure is preserved, retrieval has something real to work with. A table stays a table. A caption stays attached to its figure. A clause keeps the context of the section it lives in. The quality of every later step depends on getting this first step right.
Depth You Choose: Standard, Enhanced, Vision
Not every document needs the same depth of reading, and deeper reading costs more time and compute. Noesia makes that a choice rather than a fixed setting, so you match effort to the document in front of you.
- 1.Standard handles clean, text-based documents quickly and is the right default for most files.
- 2.Enhanced adds stronger handling of tables and layout for documents where structure carries the meaning.
- 3.Vision reads the page visually for the hard cases: scans, dense figures, and complex layouts that text extraction cannot recover.
The point is not to read everything at maximum depth. It is to read each document deeply enough that its meaning survives into retrieval.
Why It Matters Before Retrieval
Most attention in retrieval systems goes to the search and the model. But both operate on whatever the parsing step produced. Reading visually moves the investment to the earliest, most leveraged point in the pipeline: faithfully capturing what the document says before anyone asks a question about it.
Vision does not make answers louder. It makes them faithful, because the system is finally reasoning about the whole document instead of a lossy shadow of it.
You cannot retrieve meaning that was lost the moment the page was read.