How does conversational property search avoid making up listing details?
Conversational property search avoids invented details by answering only from retrieved listing records and checking each answer against them before it reaches the user. The system turns a request such as a three-bed home near good schools into structured filters plus a semantic query, retrieves candidates with hybrid keyword and vector search, and re-ranks them for fit. The language model then writes a reply that cites the listings it used. An output check compares any stated price, size or feature with the source record and blocks claims it cannot match. When nothing fits, the assistant says so and suggests close alternatives rather than stretching the truth. Octalcode built Realty One RAG for a property marketplace on this pattern: a custom RAG pipeline with hybrid retrieval, re-ranking, inline citations and output checks that block unsupported claims. Listing data is refreshed on a schedule, so sold or withdrawn homes drop out of answers.
