How does Riff prevent hallucinations or inaccurate answers from reaching buyers?

Question

Riff · Accepted Answer

TL;DR

Riff prevents hallucinations by generating answers from retrieval over your own content, not the model's general knowledge. Every response is attributed to a source document or call transcript, so buyers see where claims come from. If Riff encounters a question its knowledge base can't answer, it surfaces the gap instead of fabricating a response.

How Riff grounds answers in your content

Riff ingests your product documentation, call transcripts, sales videos, web content, and diagrams as a retrieval source. When a buyer asks a question, Riff searches that corpus first, ranks results by authority, and generates answers from the highest-confidence sources. This means responses are built from material your team knows is accurate, not from the general internet or the model's training data.

Authority ranking means curated product documentation outranks marketing pages when they disagree, ensuring buyers get the most precise version of the truth.

Attribution keeps answers verifiable

Every response from Riff carries source attribution. A buyer sees not just an answer but where it came from: "This is from your product documentation," "This is from a customer call on November 15," or "This is from your integration guide." That traceability lets buyers verify claims instantly and gives your team auditability.

If a source needs correction, you know exactly where to update it.

What happens when Riff doesn't have an answer

Riff does not invent answers to fill gaps. If a question falls outside your knowledge base, Riff surfaces that gap to your team. This creates a visible backlog of buyer questions your content doesn't yet address, which becomes your roadmap for expanding what the knowledge base covers. Over time, this feedback loop keeps the knowledge base complete and more aligned with what buyers actually ask.

The result

Hallucinations are prevented not by post-hoc filtering or confidence thresholds, but by architecture: Riff can only draw from sources it has, and it attributes every claim. Your buyers get answers anchored to your reality, not the model's guesses.

What kind of content does Riff need to get started?

Riff ingests product documentation, PDFs, slide decks, videos, call transcripts, web content, images, and diagrams. The more authoritative and current your sources, the more precise the answers.

Does Riff handle questions outside my knowledge base?

Riff surfaces knowledge gaps to your team instead of answering speculatively. You decide whether to add content, update existing sources, or refine Riff's understanding of a topic.

Can buyers see what Riff's sources are?

Yes. Every answer includes attribution, so buyers can trace claims back to the original document or conversation.

\*Verified