@@ -5,6 +5,31 @@ All notable changes to this project will be documented in this file.
55The format is based on [ Keep a Changelog] ( https://keepachangelog.com/en/1.0.0/ ) ,
66and this project adheres to [ Semantic Versioning] ( https://semver.org/spec/v2.0.0.html ) .
77
8+ ## [ 0.9.1] - 2026-08-01
9+
10+ ### Changed
11+ - ** A grounded verdict now corroborates itself with the question's own words.**
12+ ` groundedness ` tells a caller how much confidence a retrieval deserves, and it
13+ earned that from semantic similarity alone. Similarity is a strong signal for
14+ what a passage is * about* — it finds the note that answers a question phrased
15+ in words the note never uses, which is most of the value of a semantic index —
16+ but with sentence embedders the scores of same-corpus text sit in a narrow
17+ band, so it ranks candidates well and separates populations poorly. A result is
18+ now ` grounded ` when the passages are semantically close ** and** at least 60% of
19+ the question's content words are present in them. On a 24-question labelled
20+ set the verdict goes from agreeing with retrieval quality about two thirds of
21+ the time to agreeing nearly always. What no longer clears the higher bar
22+ becomes ` weak ` : the same passages are returned and the evidence is reported as
23+ thin, so nothing is withheld — only the confidence attached to it changes.
24+ Terms split on Unicode alphanumerics, so accented and non-Latin questions keep
25+ their words whole; matching is by substring, so an inflected form corroborates
26+ without a per-language stemmer; a question made only of function words abstains
27+ and leaves the decision to similarity alone.
28+
29+ A cross-encoder that scores whether a passage * answers* a question is the
30+ stronger form of the same idea. This is the part of it that needs no model, no
31+ download and no added latency.
32+
833## [ 0.8.0] - 2026-07-28
934
1035### Added
0 commit comments