Postgres handed back the catalogue, and then a second stage finished the job. The snippet extractor picked the densest cluster of matches inside that document, which belonged to an unrelated office [7]. So the model received a page that contains the sentence naming Daniel Lurie as the 46th mayor [3] together with an excerpt that does not mention him [7]. Saying it could not find anything was the correct answer to the evidence it was given [7].
The diagnostic line in the writeup is the part worth stealing. When a fact is simply missing, the gap is obvious. When a fact is unreachable instead, the system looks exactly like a model that doesn't know, as the Argus author puts it [8].
The first fix was the design most people ship. Keyword search first, vector search only when the keyword query returned zero rows [11]. That guard checks whether the search came back empty, not whether what it came back with was any good. Because the department catalogue matched almost any query with a common civic word in it, the row count was never zero, and the vector search that locates the mayor's profile at a cosine distance of 0.27 never ran at all [11][12]. The author reports that version was worse than either search used alone [11].
Running both every time and merging with reciprocal rank fusion works for a boring reason: RRF only reads each result's position in its own list, so a ts_rank and a cosine distance never have to be forced onto a shared scale [13]. It also matches the failure geometry. Keyword search demands every term appear, which is why "traffic citations" missed a document headed "Traffic Violations" [9], while vector search closes that gap but has no notion of an exact name and will rank a near-miss above the document actually about your subject [10].
The 0.42 relevance cutoff is measured rather than guessed, which I like, and measured on one corpus, which is the caveat. On-topic questions land between 0.27 and 0.34, and the two off-topic probes quoted, weather in Tokyo and renewing an Irish passport, sit at 0.51 [14]. That leaves 0.08 of headroom above the worst on-topic result and 0.09 below the off-topic reading [15]. For the number to transfer you would need the same embedding model, documents of comparable length, and an off-topic distribution sampled more widely than two questions [14].
The cost ledger carries a separate trap, and it is arithmetic. Thinking models bill reasoning tokens at the output rate but report them in a field apart from the completion, so a price computed from the two obvious numbers undercounts [19]. One measured call came to $0.000557 by the naive method against $0.001986 actually owed [20], a shortfall of $0.001429 on that single call [21].
Two publishers refused the crawler: BoardDocs 403s any client that is not a browser, and SFCTA sits behind a Cloudflare challenge, both defeatable in about ten minutes with a spoofed user-agent, which the author declined to do [16]. Those feeds are supplied by hand, pushed through the same deterministic parsers, and tagged kind='manual' [17], which is the closest thing to a staleness alarm they have [17].
In an agent shaped like this, the retrieval log is what I'd instrument first: the candidate list with both scores, plus the exact snippet that went into the prompt.
