From 1c163ead795d1b407f6053dfd6ae7296ce73c856 Mon Sep 17 00:00:00 2001 From: Zhifei Li Date: Wed, 3 Jun 2026 07:19:39 -0700 Subject: [PATCH 1/2] fix(web): restore concrete-subject search examples (bring back The Starry Night) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Search mode renders a visual grid, where a bare subject like 'The Starry Night' surfaces a striking spread of different artists' starry-night paintings (Van Gogh, Munch, Millet, the Dutch Sterrennacht). The natural-language rewrite just repeated the same painting — less compelling. Revert search examples to concrete subjects; ask mode stays conversational. --- web/components/SearchBar.tsx | 16 +++++++++------- 1 file changed, 9 insertions(+), 7 deletions(-) diff --git a/web/components/SearchBar.tsx b/web/components/SearchBar.tsx index d5867bf..c4ba9a9 100644 --- a/web/components/SearchBar.tsx +++ b/web/components/SearchBar.tsx @@ -6,14 +6,16 @@ import { Button } from "@/components/ui/button" import { Input } from "@/components/ui/input" import { getHistory, clearHistory } from "@/lib/history" -// Both modes take natural language — the embedding model is trained on it, and -// it retrieves better than bare keywords. +// Search mode shows a visual grid, so concrete subjects beat full questions: +// "The Starry Night" surfaces a striking spread of different artists' starry-night +// paintings, where a verbose question just repeats the same one. Ask mode (below) +// stays conversational. const EXAMPLE_QUERIES = [ - "What does Van Gogh's The Starry Night look like?", - "How is the periodic table laid out?", - "What does the Taj Mahal look like?", - "What is depicted in The Great Wave off Kanagawa?", - "兵马俑长什么样？", + "The Starry Night", + "Periodic table", + "Taj Mahal", + "The Great Wave off Kanagawa", + "兵马俑", ] const ASK_EXAMPLES = [ From f09a31e05115af37369bff5563df7644bf2f69be Mon Sep 17 00:00:00 2001 From: Zhifei Li Date: Wed, 3 Jun 2026 21:06:38 -0700 Subject: [PATCH 2/2] =?UTF-8?q?docs(web):=20correct=20the=20example=20comm?= =?UTF-8?q?ent=20=E2=80=94=20NL=20is=20the=20model's=20native=20query=20st?= =?UTF-8?q?yle?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The LoRA was trained on SimpleQA-style natural-language questions (naturalness- filtered), and NL queries retrieve the target page more precisely. The search grid uses bare ambiguous titles for the visual spread they fan out, not because keywords retrieve better. --- web/components/SearchBar.tsx | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/web/components/SearchBar.tsx b/web/components/SearchBar.tsx index c4ba9a9..e7857d1 100644 --- a/web/components/SearchBar.tsx +++ b/web/components/SearchBar.tsx @@ -6,10 +6,11 @@ import { Button } from "@/components/ui/button" import { Input } from "@/components/ui/input" import { getHistory, clearHistory } from "@/lib/history" -// Search mode shows a visual grid, so concrete subjects beat full questions: -// "The Starry Night" surfaces a striking spread of different artists' starry-night -// paintings, where a verbose question just repeats the same one. Ask mode (below) -// stays conversational. +// The embedding model is trained on natural-language questions (SimpleQA-style), +// which retrieve the target page most precisely — Ask mode leans on that. The +// search grid deliberately uses bare, slightly ambiguous titles instead: "The +// Starry Night" fans out to Van Gogh + Munch + Millet + the Dutch Sterrennacht, +// a far more striking visual spread than one painting repeated five times. const EXAMPLE_QUERIES = [ "The Starry Night", "Periodic table",