AI research assistant

Read papers. Get AI summaries. Cite sources. Organize notes.

Marginalia reads alongside you — turning dense PDFs into clear explanations, verifiable citations, and a linked note graph that follows you from paper to paper.

Free plan available · no credit card required

Shelves of stacked research books in a library

AI explanation

In plain terms: the model retrieves evidence before answering, which cuts hallucinated citations by 62% versus an unconstrained baseline.

Grounded · 3 sources Fig. 2, p. 4

AI summaries

Every paper distilled into plain-language takeaways — grounded in the source text, never invented.

Accurate citations

Each claim links back to its exact passage, formatted for APA, MLA, or BibTeX in one place.

Organized notes

Highlights become linked notes that carry your insights from reading into a searchable workspace.

Interactive demo · no account needed

Preprint · EMNLP 2024 Demo paper

Retrieval-Augmented Generation for Scientific Question Answering

L. Chen, M. Okafor, S. Whitfield — University of Edinburgh & Allen AI

We pair a passage-level dense retriever with a citation-constrained generator, so the model may only cite evidence it actually retrieved. Grounded-answer rates rise from 61% to 78% as retrieval depth grows to k=8, while hallucinated citations fall by 62% versus an unconstrained baseline.

Printed charts comparing experimental results — **Fig. 2** — Answer accuracy vs. retrieval depth (k). Accuracy climbs to k=8, then plateaus. Source: Chen et al., 2024.

AI Explanation

The authors replace brute-force retrieval with a small, re-ranked evidence pool — and force the generator to cite only from that pool, making every answer auditable.

Re-ranking passages beats retrieving more of them.
Citation-constrained decoding cuts invented references by 62%.
Accuracy saturates near k=8 — a practical sweet spot.

Grounded in 3 passages Confidence 0.92

Ask a question

Linked notes preview

Retrieval depth sweet spot — k=8 balances a +17-point accuracy gain against ~2.3× latency.

Method

Compare with Lewis et al. (2020) — token-level RAG vs. this paper’s passage-level re-ranking.

To read

These notes sync to your workspace once you create an account.

Plans & account

Choose your plan

Start free, upgrade whenever your reading list outgrows it. Switch or cancel anytime.

Free

For trying things out

$0 /mo

3 papers in your library
20 AI questions per month
APA & MLA citation formatting

Free plan

Pro

For serious researchers