Skip to content

test(fashion-search): NLQ vibe-expansion eval + flash-lite short-query finding - #85

Merged
octalpixel merged 1 commit into
mainfrom
feat/pipeline-integrity-s0-s7
Jun 25, 2026
Merged

test(fashion-search): NLQ vibe-expansion eval + flash-lite short-query finding#85
octalpixel merged 1 commit into
mainfrom
feat/pipeline-integrity-s0-s7

Conversation

@octalpixel

Copy link
Copy Markdown
Contributor

Adds bun run eval:vibes as a regression guard for the core 2.5.0 vibe→styles contract. Gates on realistic vibe queries (5/5 pass); reports that flash-lite degrades the NLQ on 2-word queries (verified, non-gating). No core change. 🤖 Generated with Claude Code

…ore 2.5.0)

`bun run eval:vibes` asserts cultural/mood references map to the structured `styles` attribute
(quiet luxury→minimalist+classic, old money→classic+preppy, y2k→y2k, cottagecore→romantic+bohemian,
coastal grandmother/clean girl→minimalist+classic). Lenient overlap assertion (LLM-driven).

The eval surfaced a real limitation: gemini-3.1-flash-lite degrades the WHOLE NLQ parse to nothing
on very short (2-word) queries — vibe or not (verified: "streetwear hoodie"/"cottagecore dress"/
"quiet luxury" return no parse; "blue cotton shirt" parses fine). So those fall back to the raw
embedding (no styles). Realistic 3+ word vibe queries expand reliably and are the gating contract;
the 2-word cases are marked `flaky` (reported, non-gating) until the NLQ is hardened for short input.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UeVCwYURGVrm2TFtQmWB3s
@octalpixel
octalpixel merged commit baae21e into main Jun 25, 2026
1 check passed
@octalpixel
octalpixel deleted the feat/pipeline-integrity-s0-s7 branch July 1, 2026 15:26
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant