Go back

Hallucinations and accuracy

Whether Astell's AI models hallucinate, how Astell minimizes it, and how to verify any answer.

Whether Astell's models hallucinate

Every large language model can produce a plausible-sounding statement that isn't true, and no vendor can eliminate that entirely. Astell's models are no exception. What a vendor can control is how often it happens and whether you can catch it when it does. Astell is built around both: answers are grounded in your ingested workspace data rather than the model's general knowledge, and every claim carries a citation you can check.

How Astell minimizes it

Most hallucinations happen when a model has to guess at missing context. Astell removes the guesswork before the model ever runs:

• Pre-ingested, normalized data. Your workspace is ingested ahead of time and cleaned in the pipeline: content extraction, chunking, entity resolution (the same person or project under different names becomes one identity), deduplication, and time-aware versioning so the current version outranks the stale one.
• Three stores, one shape. Every artifact lives in a relational store (structure: this is a ticket, that is an email), a vector store (meaning), and a graph (links and time). The model receives structured, connected context instead of loose search snippets.
• Permission-aware retrieval. Context is filtered to what you're allowed to see, so answers are grounded in the right slice of the workspace.
• Grounded actions. Astell Actions are drafted from the same ingested data, so proposed changes carry the right ticket, people, and history.

Verifying any answer

Every answer is sourced to the sentence that proves it. Claims link back to the original artifact (the Notion note, the Linear ticket, the Gmail thread) with the supporting passage highlighted. If something looks off, you're one click from the source. Treat the citation as part of the answer: an uncited claim is a flag, a cited one is checkable.