Every claim, cited to the line
Numbered citations link to the exact passage, page or row. Hover to preview, click to open the source.
New Prism 2: 41% fewer unsupported claims
Research that shows its work.
Prism reads your docs, tickets and the open web, then answers with sources you can check line by line. In the app for your team, or through one API.
Playground
Try an example
Connected sources
- Help desk
- Team wiki
- CRM
- Call notes
- Web
Ready
Choose an example or press Run. The answer streams in with numbered sources, the same way it does in the app and the API.
Demo answers are pre-written. In your workspace, Prism only reads what you have access to.
Trusted by research, support and product teams at 4,200 companies
Why teams switch
Built for answers you can defend in a meeting.
Most AI tools sound confident. Prism is built to be checkable: it cites, it respects permissions and it tells you when the sources don’t agree.
Permission-aware
Prism inherits access from each source. If you can’t open a file, it can’t quote it to you.
A million tokens of context
Drop in a whole data room. Prism 2 reads about 2,500 pages in one pass, no chunking strategy needed.
40+ connectors
Drives, wikis, help desks, CRMs, repos and warehouses. Synced every few minutes.
Briefs that write themselves
Schedule a question. Prism re-runs it, notes what changed and posts the brief to your team.
How it works
From question to cited answer in about two seconds.
- 01
Connect your sources
Sign in to your drive, wiki, help desk or CRM. Prism indexes in minutes and keeps every permission.
- 02
Ask in plain language
Prism plans the search, reads the relevant passages across every source and writes a draft.
- 03
Check the receipts
Each claim carries a citation. Claims Prism can’t support are flagged, not smoothed over.
Benchmarks
Fewer made-up facts. Measurably.
We test every release on three public research benchmarks and publish the harness. Higher is better; each score is the share of answers that are both correct and fully supported by their citations.
- Unsupported claims vs Prism 1.5
- −41%
- Median time to first token
- 380 ms
View as a table
| Benchmark | Prism 2 | Prism 2 Flash | Model A | Model B |
|---|---|---|---|---|
| SourceQA-Hard | 78.4 | 71.2 | 64.9 | 58.3 |
| CiteCheck 2 | 86.1 | 80.5 | 69.7 | 66.2 |
| Needle-1M recall | 91.3 | 84.0 | 72.8 | 61.5 |
Competitor models anonymised. Evaluated October 2026, temperature 0, same retrieval set for every model.
Customers
Teams that stopped double-checking everything.
“Our support leads used to spend Friday afternoons reading tickets to write the weekly summary. Prism drafts it Monday at eight, with links, and they spend twenty minutes editing.”
“We moved our in-app help search to the Prism API in a week. Deflection went from 22% to 41%, and legal signed off because every answer links to the policy it came from.”
“The security review took one call. Zero retention, SCIM and audit logs were already there. That never happens with AI vendors.”
Give your team a research assistant it can trust.
Free for individuals. Team plans start at $20 per seat with a 14-day trial.