How Prizmatic handles your data.

Who we are

Prizmatic is the evidence platform at prizmatic.ai. It runs systematic-review pipelines and related tools for people who sign in, and it serves public pages and demos to anyone. The name and address of the entity responsible for it will be added to this page before it takes effect.

What we collect and why

Five kinds of data, each for a stated reason.

• Account data. Sign-in is handled by Clerk. When you sign in we receive your Clerk user id, your name, your email address and your profile image, and we use them to show you your own reviews and nobody else's. We keep your user id, email address, display name and role in our database so that a run can be matched to the person who launched it.
• Access log. When a signed-in session touches the site we record the time, your IP address, your browser's user-agent string, the approximate location our host derives from the IP (country, region and city), your session id, your email address and your role. This is written about once an hour per signed-in session (the limiter is best-effort, so a busy session can produce a few rows an hour rather than one). The log is append-only and exists so that unauthorised access can be seen and investigated. It refuses to store anything typed into a username field that looks like a password or an API key.
• Review content. The research question, the search queries built from it, the settings you choose, the title you give a review, the records retrieved from the literature databases, the screening decisions, the extracted data, and the outputs produced from them: the study table, the evidence grid, the synthesis and the grant and aims panels. This is the work itself. We process it to produce the review and keep it so you can come back to it. The name you signed in with, the review title and the IP address your browser used to launch the run are stored alongside it.
• Demo requests. If you ask for a demo through a form on one of our landing pages, the name, email address, organisation, role and message you enter are emailed to us and used to reply.
• Analytics and advertising tags. Every page loads Google Analytics (property G-NB80TVCDVM) and a Google Ads tag (AW-17974778748). These set Google's cookies and report page views and ad conversions to Google. There is no consent banner today; the tags load for every visitor.

Who else processes it

We do not sell data. The services below process it on our behalf, each for the purpose stated.

• Clerk handles sign-in. It holds your credentials and profile and sets the cookie that keeps you signed in.
• Vercel hosts the website and runs our server code, and supplies the approximate-location headers the access log records.
• Our review backend, a server we operate, runs the pipeline and holds the job store: your research question and settings, the retrieved records, the screening decisions, the extracted data and the outputs. It receives your user id with every request so that it can refuse to show one person's review to another.
• Supabase (Postgres and file storage) holds the access log, the user record, a per-run summary (title, who ran it, status, cost, token and article counts) and copies of some run files.
• Model providers: OpenAI, Google (Gemini) and Anthropic (Claude). Your research question, the terms derived from it, your queries, and the retrieved and extracted content are sent to these providers to build queries, screen records, extract data, group findings and write the synthesis. Which stage goes to which provider is set by us and, for screening and extraction, by the model you pick in a review's settings.
• Literature services: NCBI (PubMed), Europe PMC, Elsevier (Scopus), Semantic Scholar, SerpAPI (for Google Scholar), OpenAlex and Crossref. They receive your search queries and, for citation lookups, the identifiers of the records in your review, such as DOIs and PubMed ids.
• Google receives the analytics and advertising data described above.
• Resend delivers the demo-request email.

What stays on your device

Your browser keeps its own copy of your work so that pages reopen instantly: the list of your reviews and their settings, the extraction table, the evidence-grid map, the synthesis and the tool panels. These live in your browser's localStorage and IndexedDB. They are copies of data our servers produced or you entered; the copy itself is read only by your browser. Your theme preference lives there too. Clearing site data for prizmatic.ai in your browser removes all of it, and deleting a review removes that review's copy.

How long we keep it

A retention period for review content has not been set. This is the open item the FAQ lists under data retention, and it will be filled in from a written policy rather than a guess. Until then, your reviews, the retrieved records, the extracted data and the outputs stay in the job store until you delete them.

Retention period not yet set

The access log is different. Each row is written with an expiry date, 365 days after the event by default, so that the log is purged on a schedule rather than kept forever.

Your choices

• Delete a review. Deleting a review from the dashboard removes it from this browser and, for a review that ran on our backend, sends a delete to the job store; the local copy is removed only once the backend confirms. The access log is not part of a review and is not deleted with it.
• Sign out. Signing out ends the session. The access log keeps the rows it already has.
• Clear your browser. Clearing site data removes every device copy described above.
• Ask us. You can ask what we hold about you, ask for it to be deleted, or ask a question about this page, at the contact address below.

Security

Sign-in credentials are held by Clerk, not by us. The keys our servers use to reach the backend, the model providers and the literature services never reach your browser; it talks only to prizmatic.ai and our servers attach the keys. We do not claim SOC 2 or ISO 27001 certification; the FAQ lists both as unconfirmed.

Changes to this page

This is a draft and it will change. When it takes effect the date will be at the top and we will note what changed. Until then the text describes the code as it runs, and where the two disagree the code is what happens; tell us and we will fix the page.

Contact

Contact: to be added.Open item