Screening papers without uploading them
A researcher can ask questions across up to 40 PDFs at once without sending a single file anywhere: parsing, embedding and search run in the browser tab. This page covers what the parser does with preprint PDFs, how to split a larger screen into batches, and where a small local model stops being enough.
Last updated:
The useful unit here is a folder you already curate, not a whole database. Forty PDFs at about 700 characters per chunk is a few thousand chunks, which embeds in about a minute on a laptop and stays searchable after the tab goes offline.
What the parser keeps from a preprint
| Element | What comes through | What you do about it |
|---|---|---|
| Single-column body text | In order, with page numbers on every chunk | Nothing — a citation reads “page 6” |
| Two-column layout | Usually correct, with occasional column bleed | Skim the cited passage before quoting it |
| Inline maths symbols | Partial: simple symbols survive as text | Read the derivation in the source PDF |
| Equations rendered as images | Nothing | Screenshot the equation yourself |
| Tables | Cell text in reading order, without the grid | Ask for the page and read the table in the PDF |
| Reference list | Plain text, so author names are searchable | Useful for “which paper cites X” questions |
| Supplementary CSV | Parsed row-wise, so a question can target a column | Import it alongside the PDF |
Questions a literature pass answers well
- “Which of these papers train on fewer than 10,000 examples?” — works when the number is written out in the text.
- “Where does this paper state its main limitation?” — limitation paragraphs reuse the paper’s own vocabulary, so they rank high.
- “Which papers compare against the same baseline?” — an exact-token match on the baseline name, which vector search alone tends to smear.
- “What evaluation metrics appear across this set?” — returns the metric mentions with file and page; you assemble the table.
- “Is there any mention of a replication study?” — with strictness set to documents only, an empty result is itself a usable finding.
Splitting a 200-paper screen into batches
Group by sub-topic instead of alphabetically
Forty papers that share vocabulary retrieve far better than forty unrelated titles, because a question about a baseline gets real competition among neighbours.
Name your batches in your own notes
One library lives per browser profile. Keeping “topic-a”, “topic-b” and “methods” as separate passes means a question is scoped by whatever you loaded, so track which pass you are in.
Export before you clear
Save the question list and citation sets to CSV or Markdown before clearing site data to make room for the next forty. The export is plain text and stays readable.
Re-check every claim at the PDF
Each retrieved passage carries a file name and a page number. Verify there, not at the summary paragraph.
Where a local model is not enough
- Cross-paper synthesis. The generation tier runs a 0.5B–1.5B model at 4-bit on your device. It can stitch three retrieved passages into a sentence; it cannot hold a 200-paper field in view.
- Non-text evidence. A result expressed only as a figure, an image-based equation or a colour heat map is invisible to a text pipeline.
- Scan-only material. Older scanned journals without an OCR layer produce no text, and the parser says so instead of indexing nothing.
- Mathematical verification. The tool finds where a proof is written. It does not check the proof.
When this is not the right approach
- You need to search a field, not a folder. There is no remote corpus: only the files you imported are searchable.
- Two collaborators must query the same live index. The index lives in one browser profile and is never synchronised.
- The screening set is bigger than 40 files or 200 MB and cannot be meaningfully split by topic.
- The evidence you need is in figures, scanned pages or typeset maths.
- You need a reviewable, timestamped log of what was searched. Nothing is logged, by design.
- The question demands arithmetic over many papers, such as pooling sample sizes. Retrieval returns passages; the summing is yours.
Models, offline use and what leaves the machine
- English libraries: all-MiniLM-L6-v2 at about 23 MB. Chinese libraries: bge-small-zh-v1.5 at about 25 MB. Mixed-language sets: multilingual-e5-small at about 120 MB.
- The optional generation tier downloads Qwen2.5 0.5B–1.5B at 4-bit, 400 MB to 1.0 GB, and only after you click to confirm.
- One outbound request ever: the first GET for model weights, carrying the file name. Afterwards the tool works offline.
- No account and no tracking pixels, so there is nothing to attach a half-finished review to.
- Per-library ceiling: 40 files, 25 MB each, 200 MB total, 20,000 chunks.
Can I search my whole reference manager at once?
Not through this tool. A library holds at most 40 files, so a full corpus has to be worked in batches you define.
Does it handle LaTeX source or arXiv HTML?
HTML and plain text are supported formats, so an arXiv HTML export is parsed. LaTeX source is plain text too, but macro-heavy files produce noisy chunks.
Will it extract a numbers table across papers?
It will retrieve each mention with its page, and the optional generation tier can arrange them into a list. The numbers themselves should be transcribed by you; a small local model misreads digits more often than it misreads prose.
What happens with two-column PDFs?
Reading order is usually right but not guaranteed. When a citation looks scrambled, check the passage in the PDF before quoting; the page number in the citation makes that quick.
Can I use it on a machine with no GPU?
Yes. Embedding and retrieval are CPU work. Only the optional generation tier benefits from WebGPU through WebLLM; on CPU expect roughly 3–8 tokens per second.