Install Field Report · Home Inspection RAG.
PH—015
A RAG system over regional inspection reports. The retrieval side was easy. The citation side was not. Sourced answers are the product. Here is what we shipped, where retrieval failed, and how the ontology caught it.
I What We Installed
The Home Inspection RAG install ran for a regional inspection firm with roughly forty inspectors across a mid-sized metro. The scope was a RAG system over the firm’s own inspection reports, for internal use: a senior inspector reviewing a junior inspector’s draft, looking up precedent, checking regional code interpretations.
Install ran nine days. One retrieval layer, one agent layer, one citation layer. The agent was Pattern A from Vol 05: retrieval first, agent second.
II Why the Citations Are the Product
In home inspection, an answer without a source is not an answer. Regional code varies. Material inspection standards vary. An inspector cannot take an assertion on faith. They need to see the report it came from, the section, and the year.
We made citations mandatory at the agent layer. If the agent cannot produce a citation it can point to, it refuses to answer. This was not a preference. It was a design constraint. Shipping a non-citing answer would have failed the install’s first week.
III Where Retrieval Failed
The first version of retrieval used a standard embedding model over chunked inspection reports. It worked on ninety percent of queries. It failed badly on the other ten.
The failure mode was consistent: queries that depended on regional or temporal context. “What do we typically call out on post-2020 foundation work in this zip code?” The embeddings were similarity-matching without understanding that the zip code and the date range were filters, not content.
We fixed it with hybrid retrieval: a structured filter on metadata (date, location, inspector, report type) feeding into embedding similarity on the remaining corpus. The ten-percent failure rate went to under one percent over the simulation runs.
IV How the Ontology Caught It
The home-services ontology surfaced a second issue. Some inspection language in the corpus was assertive in a way that, quoted out of context, could produce liability risk. The agent was quoting verbatim. The ontology filter flagged it.
We changed the citation format. The agent now quotes the relevant passage but also surfaces the report’s context date and inspector. The senior inspector reviewing a junior’s draft has the full provenance, not just the quote. The liability surface closes on the human, where it belongs.
The ontology did not invent this safeguard. The first home-inspection install surfaced the failure mode, and the filter rule is now native to the ontology. The next install inherited the safeguard without anyone having to ask for it.