Methodology
How our AI literature review pipeline works
What "AI literature review" means here — and what we don't do. LitReview AI is not a generic essay writer or citation database. It is a structured workflow for researchers who already have a PDF shortlist and need cross-paper synthesis with sources they can verify.
1. Bring your own papers (BYOP)
You upload up to 30 PDFs you have already screened — from Zotero, your supervisor, or discovery tools like Elicit. We do not search PubMed or invent references from an external database. Every citation in your draft must map to a file you uploaded.
2. Long-context corpus read
PDFs are parsed into full text and per-page indexes. For evidence extraction, the system sends your corpus in a single long-context pass (default ~450k characters total) using smart page-aware excerpts — Abstract, Methods, Results, Discussion, and pages matching your research question — rather than blind head truncation.
Powered by MiniMax M3 (512K–1M token context). The workspace shows estimated token use and per-PDF coverage so you know how much of each source was included.
3. Evidence table + thematic outline
The recommended flow combines evidence extraction and outline design in one pass: the model reads all PDFs, builds comparable evidence rows (findings, methods, limits, page refs), and clusters themes into an outline — Introduction, 2–4 theme sections, Research Gap, Conclusion.
Outlines are organized by themes, not paper-by-paper annotated bibliographies. You can edit every cell, choose thematic / methodological / chronological organization, and add custom evidence columns before regenerating.
4. Grounded cited draft
By default, the full draft is generated in one long-context pass over smart PDF excerpts plus your edited evidence table — all outline sections at once for cross-section continuity. If that pass cannot be parsed, the system falls back to section-by-section generation with page excerpts and rolling prior context. You remain the author: review, edit, and verify before submission.
5. Source verification
Click any citation or evidence cell to open a Context panel with a page excerpt from your PDF. This is your checkpoint against hallucination or misread findings — not a substitute for reading critical sources yourself.
6. Edit-safe synthesis sessions
AI calls reuse a multi-turn synthesis session so your corpus is loaded once and later steps (outline, draft) build on the same context — with MiniMax M3 prefix caching where supported.
When you edit the evidence table or outline, stale AI context is cleared automatically. Regeneration uses your latest edits; corpus excerpts are only re-sent when needed, with visible coverage stats in the workspace.
What we do not claim
- Literature discovery or PRISMA-scale systematic reviews
- Scite-style supporting / contrasting citation graphs
- Automatic guarantee of factual accuracy — always verify against originals
- Replacement for your judgment, institution policies, or authorship responsibility
Typical stack
- Discover papers — Elicit, Consensus, Research Rabbit
- Manage library — Zotero (+ optional .bib import here)
- Synthesize & draft — LitReview AI (this tool)
- Final edit — Word / thesis template