Help Center

Help & support

Everything you need to run your first review, from install to export.

New to Compendium? See how it works has the full walkthrough. Free beta right now, contact us to get started. First launch walks you through choosing and downloading the AI model, with an explicit consent step before anything downloads.

Before you start: Compendium is free beta software, provided AS IS, with no warranty of scientific accuracy. It's a research assistant, not a finished product: you're solely responsible for every scientific decision. The AI abstains rather than guesses and never decides inclusion or exclusion on its own. Always verify results against primary sources. Your research data stays on your device.

System requirements

The AI runs on your machine, so hardware matters. Figures are approximate.

Component Minimum Recommended High performance
RAM8 GB16 GB16 GB+
GPUNot required8 GB VRAM (optional)24 GB VRAM
Disk (app + model)~4 GB~8 GB~24 GB
AI model3B-class (e.g. Llama 3.2 3B)7–8B-class (e.g. Qwen3 7B)32B-class
Best forScreening, PICO draftingThe full intended experienceManuscript drafting, complex extraction

Windows 10+, macOS 12+, recent Linux. Internet needed once, for the first model download.

Older hardware? Compendium auto-suggests a lighter model. Still slow? Switch to a smaller model in Settings → AI, or contact us. On a 3B model, pilot 20–30 references before trusting it fully.

Still need help?

Email support

Any issue you can't resolve on your own.

askcompendium@gmail.com
Reply within 2 business days.

Found a bug?

Use "Report an issue" in the app. It sends version and description, no research data attached.

Goes straight to the development queue.

Practical questions

Which file formats can I import?

RIS and CSV, the standard exports from Embase, Cochrane, Web of Science, Scopus, Mendeley, Zotero and EndNote. PubMed's native NBIB export is also supported directly, no conversion needed.

Can I work offline?

Yes, once the model is downloaded. Screening, extraction, analysis and the license check all run locally. There is no cloud option and nothing to log into.

Do I need to already know PRISMA or other reporting standards?

No. Compendium explains each guideline item as you go, generates the PRISMA 2020 flow diagram from your numbers, and flags anything unaddressed before export.

Is the AI output trustworthy enough to cite or publish?

AI suggestions are assistance, not conclusions. Every action is logged separately from your final call. Treat them like a junior reviewer's first pass: a starting point you verify.

Troubleshooting

The AI model won't load / "AI unavailable"

Usually the model didn't finish downloading, or the AI engine failed to start.

  1. Check you have at least 3 GB of free disk space.
  2. Settings → AI → Model status. If "Incomplete," click Re-download.
  3. If "Not running," click Restart AI engine and wait 30 seconds.
  4. On Windows, confirm your antivirus hasn't blocked the AI process.
  5. Still stuck? Contact support at askcompendium@gmail.com with your OS and the error message.
Import failed / references didn't appear
  1. Confirm the file is a valid RIS or CSV.
  2. PubMed exports as NBIB, not RIS. Bring it into Zotero, Mendeley or EndNote first and export as RIS.
  3. For CSV, make sure it has column headers (at minimum Title, Abstract, Year).
  4. Still failing? Re-export as RIS from Zotero or Mendeley, which format it more reliably.
The app is slow or freezes during screening

Most often the model is too large for the available RAM.

  1. Settings → AI → Current model. On 8 GB RAM running a 7B model, switch to a 3B model.
  2. Close other memory-heavy apps while screening.
  3. Still slow? Contact support if the 3B model is still too slow for your machine.
The meta-analysis or forest plot looks wrong
  1. Confirm every included study has a complete row; rows with missing data are excluded and flagged.
  2. Check the effect measure matches your data (OR/RR for binary; MD/SMD for continuous; HR for time-to-event).
  3. A meta-analysis needs at least 2 studies; with 1, a narrative synthesis is the right output.
  4. If numbers differ from a reference tool, export the study data as CSV and send it to support with the measure and model used.
The PRISMA diagram shows wrong numbers
  1. "Records identified" is the total imported before deduplication.
  2. "Duplicates removed" reflects what you confirmed as duplicates; skipping that step shows 0.
  3. Exclusion counts follow your Screening decisions. Click any value in the diagram to override it before export.