Case study by NerdWare Tech. The demo runs on sample pages; numbers marked illustrative are not measurements.

Search the page, not the text.

We embed PDF pages as images, so the tables and charts that text extraction destroys stay answerable.

Interactive retrieval demo

  1. Coarse pooled search1 vector per page
  2. MaxSim rerankfull multivector
  3. VLM answerwith bbox citation

Light table

6 pages indexed

    coarse MaxSim (illustrative scores)

    Page viewer

    Run a question to light up pages.

    Cited answer

    illustrative

    The VLM only sees the top pages, as images, and has to point at the region it used.

    One page in, one cited region out.

    Five stages. Select one to see what it hands to the next.

    MaxSim, one query token at a time.

    Every query token finds its best-matching patch on the page. The page score is the sum of those maxima.

    Page
    Query token (q​i)

    Score(Q,D) = Σi maxj ( Eqi · EdjT )

    Why not chat with your PDF.

    Chat-with-PDF flattens a page into text before it ever searches. Here is the same chart both ways.

    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.

    Embed the page as an image

    Bar height, colour and axis stay together. The answer points at a box.

    Scored on retrieval, before any answer.

    ViDoRe-style evaluation against a dense text bi-encoder. These bars are illustrative, not results.

    Metric

    TARGET, ILLUSTRATIVE

    Latency budget per query (target)

    Stack.

    Have documents your search can’t read?

    Tell us what you need answered from them. The demo above uses sample pages.

    Book a free call