feat(recall): 检索为王——op=search 模糊搜索 + days/day 按日期翻页,退役盲列 list

用户定调:记起的往往是「一段事」而非「第几条」,search 价值高于 list;list 也必须按日期走而非盲目罗列。
recall 从 list/files/link 升级为 search/days/day/files/link:

- op=search {query, count}:碎片级全库模糊检索(主入口)。空白/标点分词多词择优,
  子串直匹;CJK 长词整串不中时二元字组过半即模糊命中(得分 0.9 折算),英文大小写不敏感;
  结果按相关度排序、标注匹配百分比与所属日期,默认 10 条上限 30;无命中回友好提示并引导 days。
- op=days:日期索引(哪天、几条)——只记得「大概哪天」时的入口;op=day {date} 取整天完整日记。
- 开场注入头行与工具 description 同步改口径(提起旧事 → search;记时间 → days/day);
  存储/存在性判定/files/link 不变。op=list 退役(含 skip 盲翻页参数)。

探针 12 项全过:精确/模糊(打折分)/多词排名/大小写/无命中/days 计数/day 整天/空天/日期校验/退役提示。
This commit is contained in:
2026-09-29 14:52:12 +08:00
parent c7fba5006e
commit 0307161e6f
3 changed files with 81 additions and 31 deletions
+50 -20
View File
@@ -87,37 +87,66 @@ function journalFor(db, cfg, user) {
groups[groups.length - 1].frags.unshift(p.frag);
}
const head = '【长期记忆·日记(最近' + picked.length + '条)】收工时的当日事件摘要,近日的在前、天内按发生顺序;'
+ '每条【编号】是来源会话 id:更早的日记用 recall 工具 op=list 翻,那次工作的成果文件用 op=files 查、op=link 取。';
+ '每条【编号】是来源会话 id:更早的日记用 recall 工具 op=search 关键词检索、op=days 按日期翻页再 op=day 取整天;'
+ '那次工作的成果文件用 op=files 查、op=link 取。';
const parts = [head];
for (const g of groups) parts.push('=== ' + g.day + ' ===\n' + g.frags.join('\n\n'));
return parts.join('\n\n');
}
// ---------- 读:回忆更早碎片(recall 工具 op=list 的数据源) ----------
// 与 journalFor 同展开顺序(全局新在前);skip 默认跳过开场注入窗,count 默认 20 上限 50。
// 尾部标注剩余更早条数与续翻参数,AI 读完自然知道怎么翻页。
function olderFragments(db, cfg, user, skip, count) {
// ---------- 读:关键词检索(recall 主入口:用户记起的是「一段事」不是「第几条」) ----------
// 碎片级全库搜索:存储量小(每天一行、内容几十行),内存打分足够,不引入 FTS。
// 分词:空白/标点切词(多词命中多者为先);单字命中 = 子串包含;CJK 长词未整串命中时
// 按二元字组兼容(过半才算模糊命中,得分打折)——支撑「记个大概」的模糊查。英文词大小写不敏感。
// 无命中返回 ''(友好提示由调用方给)。
const CJK_RE = /[\u4e00-\u9fff]/;
function tokenScore(fl, t) {
if (fl.includes(t)) return 1;
if (t.length < 2 || !CJK_RE.test(t)) return 0;
let hit = 0;
for (let i = 0; i + 2 <= t.length; i++) if (fl.includes(t.slice(i, i + 2))) hit++;
const frac = hit / (t.length - 1);
return frac >= 0.5 ? frac * 0.9 : 0;
}
function searchFragments(db, user, query, count) {
const toks = String(query || '').toLowerCase().split(/[\s,,、;;。.!!??]+/).filter(Boolean).slice(0, 8);
if (!toks.length) return '';
const lim = Math.max(1, Math.min(30, Number(count) || 10));
const rows = db.prepare('SELECT day, content FROM memory_entries WHERE user = ? ORDER BY day DESC').all(String(user || ''));
const flat = [];
const hits = [];
let seq = 0;
for (const r of rows) {
const frags = String(r.content || '').split('\n\n').filter(Boolean);
for (let i = frags.length - 1; i >= 0; i--) flat.push({ day: r.day, frag: frags[i] });
for (let i = frags.length - 1; i >= 0; i--) { // 天内新在前,seq 自然全局新在前
const frag = frags[i];
seq++;
const fl = frag.toLowerCase();
let sum = 0;
for (const t of toks) sum += tokenScore(fl, t);
if (sum > 0) hits.push({ day: r.day, frag: frag, score: sum / toks.length, seq: seq });
}
}
const skipN = Number.isFinite(Number(skip)) ? Math.max(0, Math.floor(Number(skip))) : injectMax(cfg);
const lim = Math.max(1, Math.min(50, Number(count) || 20));
const page = flat.slice(skipN, skipN + lim);
if (!page.length) return '';
const groups = [];
for (const p of page) {
if (!groups.length || groups[groups.length - 1].day !== p.day) groups.push({ day: p.day, frags: [] });
groups[groups.length - 1].frags.unshift(p.frag);
}
const rest = flat.length - skipN - page.length;
const parts = ['【更早的日记】第 ' + (skipN + 1) + ' 条起共 ' + page.length + ' 条' + (rest > 0 ? ',继续翻:skip=' + (skipN + page.length) : ',这已是全部记忆')];
for (const g of groups) parts.push('=== ' + g.day + ' ===\n' + g.frags.join('\n\n'));
if (!hits.length) return '';
hits.sort((a, b) => b.score - a.score || a.seq - b.seq);
const shown = hits.slice(0, lim);
const parts = ['【日记检索】「' + String(query).trim() + '」命中 ' + hits.length + ' 条,按相关度列前 ' + shown.length + ' 条'
+ (hits.length > shown.length ? ';要更多加大 count 或关键词收窄' : '') + ':'];
for (const h of shown) parts.push('--- ' + h.day + '(' + Math.round(h.score * 100) + '% 匹配)---\n' + h.frag);
return parts.join('\n\n');
}
// ---------- 读:日期索引(只记得「大概哪天」时,先看有哪些天有账) ----------
function daysIndex(db, user) {
const rows = db.prepare('SELECT day, content FROM memory_entries WHERE user = ? ORDER BY day DESC').all(String(user || ''));
return rows.map(r => '- ' + r.day + '(' + String(r.content || '').split('\n\n').filter(Boolean).length + ' 条)');
}
// ---------- 读:某天的完整日记 ----------
function dayEntry(db, user, date) {
const r = db.prepare('SELECT content FROM memory_entries WHERE user = ? AND day = ?').get(String(user || ''), String(date || '').trim());
return r ? String(r.content) : '';
}
// ---------- 裁剪:day 为 ISO 串,字典序即日期序;默认永久存储为 no-op ----------
function prune(db, cfg) {
const cutoff = cutoffDay(cfg);
@@ -128,6 +157,7 @@ function prune(db, cfg) {
module.exports = {
retainDays, injectMax, dayString, cutoffDay,
appendFragment, markJournaled, journalFor, olderFragments, prune,
appendFragment, markJournaled, journalFor,
searchFragments, daysIndex, dayEntry, prune,
INJECT_CAP, DAY_MAX, FRAG_MAX,
};
+13 -7
View File
@@ -235,7 +235,7 @@
"type": "function",
"function": {
"name": "recall",
"description": "Access your long-term memory: the work diary. Each finished session leaves one short fragment (a few dozen characters, headed by the source session id), stored permanently; only the newest fragments were injected at session start. Read-only.\n\nop=list Older diary fragments beyond the injected window. Optional: skip (default = skip the injected window), count (default 20, max 50).\nop=files List files of a past session that still exist on disk right now. Requires id (the session id from a diary fragment's 【…】 heading).\nop=link Mint a temporary download link for one of those files. Requires id and path (a name from op=files).",
"description": "Access your long-term memory: the work diary. Each finished session leaves one short fragment (a few dozen characters, headed by the source session id), stored permanently; only the newest fragments were injected at session start. When the user vaguely recalls something from the past, search is your primary entry — you cannot know what you have forgotten. Read-only.\n\nop=search Requires query (keywords; space-separated words rank better, e.g. '身份证 水印'). Fuzzy fragment-level match across ALL dates (Chinese tolerant to partial phrases), best matches first. Optional count (default 10, max 30).\nop=days No arguments. Index of diary days (date + fragment count) — use when the user remembers roughly WHEN but not what; then op=day.\nop=day Requires date (YYYY-MM-DD, as listed by op=days). Full fragments of that day.\nop=files Requires id (the session id from a diary fragment's 【…】 heading). Files of that past session that still exist on disk right now.\nop=link Requires id and path (a name from op=files). Mint a temporary download link.",
"parameters": {
"type": "object",
"required": [
@@ -245,19 +245,25 @@
"op": {
"type": "string",
"enum": [
"list",
"search",
"days",
"day",
"files",
"link"
],
"description": "What to recall: list = older diary fragments; files = a past session's current artifacts; link = a download link for one of them."
"description": "What to recall: search = keyword fuzzy over all diary fragments (primary); days = which dates have entries; day = one date's full diary; files = a past session's current artifacts; link = a download link for one of them."
},
"skip": {
"type": "integer",
"description": "op=list: how many newest fragments to skip. Default = the window already injected at session start."
"query": {
"type": "string",
"description": "op=search: keywords to look for (space-separated words rank best, e.g. '身份证 水印'). Chinese and English both supported, case-insensitive."
},
"count": {
"type": "integer",
"description": "op=list: fragments to return (default 20, max 50)."
"description": "op=search: max fragments to return (default 10, max 30)."
},
"date": {
"type": "string",
"description": "op=day: the day to open, YYYY-MM-DD exactly as shown by op=days."
},
"id": {
"type": "string",
+18 -4
View File
@@ -1226,9 +1226,23 @@ function toolRecall(ctx, args) {
return { sid: sid, title: s.title || sid, dir: path.join(root, sid) };
};
switch (op) {
case 'list': {
const txt = memory.olderFragments(db, ctx.cfg, user, args.skip, args.count);
return txt || '更早的日记没有了:开场注入的就是全部记忆。需要某次工作的成果文件用 op=files。';
case 'search': {
const q = String(args.query || '').trim();
if (!q) throw new Error('op=search 需要 query(关键词;空格分隔多个词更好,如「身份证 水印」)');
const txt = memory.searchFragments(db, user, q, args.count);
return txt || '日记库里没有匹配「' + q + '」的条目。换个更宽泛的关键词再试,或 op=days 看哪天有什么;成果文件走 op=files。';
}
case 'days': {
const list = memory.daysIndex(db, user);
if (!list.length) return '日记库是空的:还没有任何收口条目(开场注入的日记也出自这里)。';
return '【日记日期索引】共 ' + list.length + ' 天有记录(新在前):\n' + list.join('\n')
+ '\n用 op=day {date} 看某天的完整日记;记得关键词就直接 op=search。';
}
case 'day': {
const date = String(args.date || '').trim();
if (!/^\d{4}-\d{2}-\d{2}$/.test(date)) throw new Error('op=day 需要 date(YYYY-MM-DD,先用 op=days 查有哪些天)');
const txt = memory.dayEntry(db, user, date);
return txt ? date + ' 的日记:\n' + txt : date + ' 没有日记(当天无交互或条目已裁剪;op=days 可看有哪些天)。';
}
case 'files': {
const t = resolveSession();
@@ -1260,7 +1274,7 @@ function toolRecall(ctx, args) {
return pl.base_url + '/p/' + publicFiles.mint({ sessionId: t.sid, rel: rel, ttlMs: pl.ttl_ms });
}
default:
throw new Error('未知的 op:"' + op + '"。可用:list(翻更早日记,可选 skip/count)/ files(需 id,列那次会话现存成果文件)/ link(需 id+path,产临时下载链接)');
throw new Error('未知的 op:"' + op + '"。可用:search(需 query,可选 count)/ days(日期索引)/ day(需 date)/ files(需 id)/ link(需 id+path)');
}
}