fix: web_search에 검색실패 프로토콜 규칙 — web_fetch 우선·쿼리 재구성·기억 답변 금지
09-24 채팅 실측 2라운드: 쿼리 분리 규칙은 적용됐으나(콤마 나열 소멸 확인)
①모델이 잡동사니 1위 결과만 보고 web_fetch를 안 쓰고 실패 선언 — 같은 목록에
정답 페이지("2MP Mira220 USB3 Evaluation Kit - Arducam Wiki")가 있었음
②일반어("resolution specifications") 붙은 쿼리가 쇼핑/매뉴얼 페이지만
돌려오는데 "<product> datasheet" 재구성이면 datasheet 소스가 나옴(직접 실측)
③실패 선언 후 기억으로 스펙 표 작성 → 1.92MP 환각(실측 1456×1088≈1.58MP),
같은 대화에서 모델 자신의 정답과 모순.
Co-Authored-By: Claude Code <noreply@anthropic.com>
This commit is contained in:
@@ -164,7 +164,7 @@ export function createBuildTools(isOrchestrationSkillEnabled: () => boolean) {
|
||||
type: 'function',
|
||||
function: {
|
||||
name: 'web_search',
|
||||
description: 'Search the web for current information. Use web_fetch on result URLs to read full page content. For a SPEC COMPARISON between two specific things (e.g. "RTX 3090 vs RTX 4080 Super memory bandwidth"), do NOT combine them into one "A vs B" query — search for each one\'s spec separately ("RTX 3090 memory bandwidth GB/s", then "RTX 4080 Super memory bandwidth GB/s"). Verified 2026-08-10: a combined comparison query returned only a generic overview page with neither figure, while the same two searched separately immediately returned the real numbers for both (936 GB/s vs 736 GB/s). "vs"/comparison phrasing tends to surface review/overview pages instead of the specific numbers a spec question needs. Same rule for comma-separated product lists: never pack several product/model names into one query ("Arducam IMX296 resolution, Basler dart M resolution, Mira220 resolution") — engines return unrelated junk for it (verified 2026-09-24: three consecutive comma-list queries returned only junk pages while the same subjects searched one per query immediately returned spec pages). One product per query.',
|
||||
description: 'Search the web for current information. Use web_fetch on result URLs to read full page content. For a SPEC COMPARISON between two specific things (e.g. "RTX 3090 vs RTX 4080 Super memory bandwidth"), do NOT combine them into one "A vs B" query — search for each one\'s spec separately ("RTX 3090 memory bandwidth GB/s", then "RTX 4080 Super memory bandwidth GB/s"). Verified 2026-08-10: a combined comparison query returned only a generic overview page with neither figure, while the same two searched separately immediately returned the real numbers for both (936 GB/s vs 736 GB/s). "vs"/comparison phrasing tends to surface review/overview pages instead of the specific numbers a spec question needs. Same rule for comma-separated product lists: never pack several product/model names into one query ("Arducam IMX296 resolution, Basler dart M resolution, Mira220 resolution") — engines return unrelated junk for it (verified 2026-09-24: three consecutive comma-list queries returned only junk pages while the same subjects searched one per query immediately returned spec pages). One product per query. If the top results look irrelevant (YouTube help pages, shopping sites, community pages), do NOT declare failure and do NOT answer spec numbers from memory — two steps instead: (1) web_fetch the most promising result URL anyway; the useful page is often in the list even when result #1 is junk (verified 2026-09-24: "Arducam Mira220 product page" returned the correct Arducam Wiki evaluation-kit page while other queries around it returned junk — reading that one page would have answered the question). (2) Rephrase: drop generic spec words like "resolution specifications" and try "<product> datasheet" or the bare product name (verified 2026-09-24: "Arducam IMX296 resolution specifications" returned only shopping/manual pages, while "IMX296 datasheet" returned actual datasheet sources). Never fill a spec table from memory after failed searches — that produced a wrong 1.92MP figure for the IMX296 (real: 1456×1088 ≈ 1.58MP), contradicting the model\'s own earlier correct answer in the same conversation.',
|
||||
parameters: {
|
||||
type: 'object', required: ['query'],
|
||||
properties: { query: { type: 'string', description: 'Search query' } },
|
||||
|
||||
Reference in New Issue
Block a user