The engineering behind the answers
A deeper read for the engineers, security reviewers, and operators who want to understand what's happening under the hood. No black boxes.
Four steps from question to answer
- Step 1
Search the entire web or bring your sources
Search the open web — government data, journalist reporting, academic papers, YouTube — or upload your own files and URLs. Each retrieval is scoped to what you chose; nothing outside that scope gets quoted.
- Step 2
Generate with quote constraints
The model is instructed to produce verbatim quotes from the retrieved passages. Generation leans on the retrieved text rather than free-form output.
- Step 3
Verify deterministically
Each generated quote is checked against the cited passage using a sliding-window match with a high similarity threshold. Mismatches are removed; the model retries against what it actually has.
- Step 4
Surface evidence inline
Verified quotes render with the source one click away. Every claim is reachable back to the passage it came from.
An agent that's mostly code
The four steps above are deterministic code, not a prompt asking a model to behave. The language model is one component, invoked only where generation genuinely needs it — retrieval scoping, quote verification, and refusal are ordinary code with predictable behavior. That's what lets the same question return the same evidence, and what makes the pipeline auditable end to end.
Why deterministic verification matters
Statistical confidence scores tell you how sure a model is, which isn't the same as whether it's right. A deterministic string match against the source tells you whether the quote actually appears. That's the standard reviewers actually need.
What happens when nothing supports the answer
If the model can't anchor a quote to a retrieved passage, the claim is removed before the answer reaches the user. When nothing in the source supports the question, the answer is a plain "not found" rather than a fabrication.
Security and privacy posture
Documents are scoped to the account that uploaded them, and we don't train frontier models on customer content. For specifics — residency, retention, attestations, training and internal-use carve-outs — talk to us.
How it works FAQ
- What model is behind this?
- Cemented AI uses third-party foundation models with our own prompting and retrieval pipeline. We treat the model as a component, not the product — the product is the verification and evidence layer around it.
- Can it hallucinate inside summary text?
- Free-form prose can still be imprecise. Anything inside quotation marks is verified; the discipline we encourage is to lean on quotes rather than summaries when accuracy matters.
- What if I already trust my AI vendor?
- Then a single side-by-side test on a document you know cold should be enough to see the difference. We'll set it up.
- What does it cost?
- Free for individuals, Pro and Max for teams. Enterprise covers single sign-on, tighter data-handling, and custom integrations — write to [email protected].
- How do I get started?
- Sign in and ask a question. Or schedule a deeper technical session.
Run a side-by-side test
Sign in to try it, or schedule a technical walkthrough.