Fusion of body, vector and prose signals, not just the word "refunds," carries billing_credit_note into the top six

Rank fusion buys the fourth channel cheaply because K is large. First place pays 1/61, which is 0.01639. Twenty-second pays 1/82, or 0.01220 ^[13][14]. A hit at rank 22 is therefore worth about 74 percent of a hit at rank 1 ^[15]. The embedding channel put twenty-one tables ahead of the right one and still supplied roughly 27 percent of the winning total ^[16][8]. Between the body channel at first and the prose channel at second, the difference was 0.00026 ^[17].

What the prose channel matched on is one generated sentence. The description reads "Records refunds or corrections issued against a specific invoice." ^[6] "refunds" is the term that links "gave back" to a credit note, and it appears nowhere in the DDL ^[7]. The names channel, which reaches 100 percent recall@6 when people phrase questions in the schema's own vocabulary, matched nothing at all here ^[4][3]. The post also notes there is no per-term score to inspect when a dense ranking comes back slightly off ^[9].

Those four rankings are one question on a 42-table schema, measured on a single run ^[2][33]. The wider evidence is a 52-question fixture set, and the author says he got its central question wrong first, at the cost of a public retraction ^[21]. One flat bag over names, columns, types, comments and descriptions answered 41 of 52, or 78.8 percent; giving prose its own index answered 48, or 92.3 percent ^[22][23]. That is a gap of seven questions ^[24]. Inside one bag the sentences compete with the identifiers for term frequency, and the identifiers lose, because a description has more words in it than a name does ^[25].

In the names index, "contact" appears in 17 of 1,245 objects, and its IDF is 4.27 ^[27]. Across the generated descriptions it appears in roughly 1,072 of 1,245, about 86 percent of the corpus, and IDF falls to 0.15 ^[26][28][27]. An accurate English word, generated into 86 percent of the documents, then scores like "the". The post says the descriptions were accurate and that their accuracy does not save them, because IDF is computed over the corpus and not over the sentence ^[30].

The cost of ranking-based fusion lands on tie-breaks. A channel that ordered its results for no meaningful reason still votes at full strength, so if every description contains the word "about", every object ties in the prose channel and the order the catalog was built in decides the ranking ^[18]. Shuffling insertion order changed which tables came back on 1 to 2 questions out of 12, roughly 8 to 17 percent, without moving recall ^[19][32]. A commenter on the author's previous post raised this, and the author says both fixes proposed are right and both are filed ^[20].

For the shipped 94.2 percent to transfer, two conditions have to hold in your catalog. Users have to ask in business vocabulary, because on schema-vocabulary questions the names channel already returns the answer in the top six on its own ^[4][22]. And the generated descriptions have to avoid putting one common word into most of the corpus, which is what happened at 1,245 objects ^[26]. The post credits fielded scoring for the fix, and says the reason is structural, not a tuning choice ^[31].