Flat retrieval
BERT achieved around 93–94% on SST-2.
- No section awareness — the number could be from related work
- No source you can check
- Nothing to paste into a literature review
Retrieval for research papers
Most tools read a paper as one flat page. This one reads it as a paper.
Sections, authors, venue, year, citations — parsed on the way in, then used at the moment you ask. Every passage it quotes comes back with the section it came from and a citation you can check.
Runs on local models. Your papers never leave the machine.
Journal of Big Data
(2025) 12:84
A survey of reinforcement learning for coordination in traffic light control
3. Methodology
Each agent observes queue length and phase duration at its own intersection. We train a multi-agent deep Q-network with a shared replay buffer, optimising with Adam at a learning rate of 0.0003 for 500 episodes.
Coordination emerges without a central controller (Saadi et al., 2025).
“How was it trained?”
A multi-agent deep Q-network with a shared replay buffer, Adam at 0.0003 for 500 episodes (Saadi et al., 2025, methodology)
§ 01 — The difference
The same question, asked two ways
Generic retrieval chunks a PDF and hopes the right paragraph floats up. It usually gets close. Close is not a citation.
DocuSense
BERT achieved 93.5% F1 on SST-2 (Devlin et al., 2018, results)
- Traced to the results section, so you can see where the number came from
- Citation checked against the source before it is shown
- Reference list and BibTeX come with it
§ 02 — What it knows
Structure, not just text
It finds the paper inside the PDF
Title, authors, venue, year, DOI, abstract, twenty section types, and citations in both numbered and author–year form — with a confidence score, so a lecture handout is not treated as a journal article.
It tells you which section an answer came from
Every chunk is tagged with its section, read from the chain of headings above it, so a passage under “Experiments › Baselines” is filed as experiments. Answers cite the section alongside the author and year. Restricting a search to one section is available too, and ships off: measured on QASPER it hid the answer more often than it found it.
It reads your filters out of plain English
“Papers from 2020–2023 by Bengio” becomes a year range and an author filter applied inside the vector search, before ranking — not bolted on afterwards where it would quietly break the top results.
It keeps exact terms exact
Embeddings think BERT-base and RoBERTa are neighbours. BM25 knows they are not. Both rankings are fused, so meaning and precision survive together.
It deletes citations it cannot prove
Small models invent plausible references even when told not to. Every citation is checked against the retrieved sources by author and year. Unsupported ones are removed rather than shown to you.
§ 03 — Method
Three steps, no configuration
-
1
Drop in a PDF
It is converted, cleaned, parsed for structure, split along section boundaries, embedded, and indexed. A 30-page survey takes about half a minute.
-
2
Ask in your own words
Your phrasing decides where it looks and what it filters on. Follow-up questions inherit the thread, so “what about its accuracy?” resolves against whatever you were just discussing.
-
3
Read the answer with its receipts
Text arrives as it is written, each claim tied to the chunk it came from, with an APA reference list and BibTeX export underneath.
Open a shelf
Your papers, actually searchable.
Free, local, and private. Nothing is uploaded anywhere you did not put it.
Create an account