Your papers have figures, tables, and equations. Most PDF assistants read past them.
GlyphScholar indexes both the words on the page and the page itself, so an answer can point to a diagram or a formula — not just paraphrase the sentence next to it.
Fig. 2 — throughput vs. batch size
Throughput peaks around a batch size of 64, then falls off.
p. 4 · Fig. 2How a document becomes something you can ask questions of
Upload a PDF
MinerU's layout parser reads the document the way a person would — separating running text from figures, tables, and equations instead of flattening everything into one text stream.
It's indexed twice
Paragraphs go into a hybrid text index — dense embeddings re-ranked with BM25. Every figure, table, and equation is cropped and embedded as an image in a separate visual index, PixelRAG.
Ask, and get a grounded answer
A question is matched against both indexes at once. If the answer lives in a chart rather than a caption, that's what gets retrieved — and the reply points back to the exact page it came from.
What makes the retrieval different
- Hybrid text retrieval
- Dense pgvector similarity and BM25 keyword re-ranking run together, so a question can be phrased loosely or match a term in the document exactly and still land on the right passage.
- PixelRAG visual grounding
- Figures, tables, and equations are embedded as images, not summarized into a caption first. A question about a trend line or a formula is matched against the pixels, not a lossy description of them.
- Page-level citations
- Every answer carries a reference back to the page and region it was drawn from, so you can check a number or a claim against the source in one click instead of taking it on faith.