From e1fd68e8cbb668112806f663c919fdbe2c420df2 Mon Sep 17 00:00:00 2001 From: Yiorgis Gozadinos Date: Sat, 22 Aug 2026 12:56:20 +0300 Subject: [PATCH] Say that searching several databases wants a reranker Reciprocal rank fusion compares ranks, so every database contributes its own best matches whether or not they answer the question, and results from databases holding nothing relevant displace better ones. Measured on one corpus split three ways over 3,045 queries: retrieval MAP 0.6044 without a reranker against 0.9798 for the same corpus in a single database, and 0.9918 against 0.9914 with one. --- docs/configuration/storage.md | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/docs/configuration/storage.md b/docs/configuration/storage.md index 6212ccd5..6a8a52b7 100644 --- a/docs/configuration/storage.md +++ b/docs/configuration/storage.md @@ -213,6 +213,15 @@ reranker where there is one and by reciprocal rank fusion otherwise. Each result carries `source`, the name of the database it came from, and so does each citation. +**Configure a reranker when searching several databases.** Reciprocal rank fusion +compares ranks, not scores, so every database contributes its own best matches +whether or not they are relevant to the question, and results from databases +holding nothing relevant displace better ones. A reranker scores the whole union +instead, which removes the effect. Measured on one corpus split three ways, with +the same queries: retrieval MAP 0.9914 with a reranker against 0.9918 for the +same corpus in a single database, and 0.6044 without one against 0.9798. The cost +is that a reranker scores candidates in proportion to the number of databases. + A database that cannot be opened fails the whole query and is named in the error. A result set silently missing one of the databases asked for cannot be told apart from a complete one.