Extract, chunk, embed text
Revenue by region Figure 2. Net revenue per quarter, USD millions (synthetic data) EMEA APAC AMER 60 40 20 0 Q1 Q2 Q3 Q4 52
The bars are gone. A lone 52 floats with no region and no quarter.
Case study by NerdWare Tech. The demo runs on sample pages; numbers marked illustrative are not measurements.
We embed PDF pages as images, so the tables and charts that text extraction destroys stay answerable.
coarse MaxSim (illustrative scores)
The VLM only sees the top pages, as images, and has to point at the region it used.
Five stages. Select one to see what it hands to the next.
Every query token finds its best-matching patch on the page. The page score is the sum of those maxima.
Score(Q,D) = Σi maxj ( Eqi · EdjT )
Chat-with-PDF flattens a page into text before it ever searches. Here is the same chart both ways.
Revenue by region Figure 2. Net revenue per quarter, USD millions (synthetic data) EMEA APAC AMER 60 40 20 0 Q1 Q2 Q3 Q4 52
The bars are gone. A lone 52 floats with no region and no quarter.
Bar height, colour and axis stay together. The answer points at a box.
Rows and columns keep their geometry. No parser guesses where a cell ends.
The answer returns a page and a normalised box, so a reviewer looks at the evidence, not a chunk id.
One model reads layout, figures and text. Fewer moving parts to fail silently.
ViDoRe gives page-level relevance labels, so retrieval is scored before any LLM is involved.
ViDoRe-style evaluation against a dense text bi-encoder. These bars are illustrative, not results.
TARGET, ILLUSTRATIVE
Tell us what you need answered from them. The demo above uses sample pages.
Book a free call