tt-help-cli-ycl 1.3.93 → 1.3.95

package/src/watch/tag-service.js

@@ -79,9 +79,66 @@ export function parseTagsFromResponse(content) {
   return tags;
 }
 
+/**
+ * 解析 LLM 的两阶段响应（strategy + tags）
+ * 期望格式: {"strategy": "...", "tags": [...]}
+ * 兼容旧格式: 纯数组或纯文本
+ */
+export function parseDiscoverResponse(content) {
+  let strategy = null;
+  let tags = [];
+
+  try {
+    const parsed = JSON.parse(content);
+
+    // 新格式: { strategy, tags }
+    if (parsed.strategy && Array.isArray(parsed.tags)) {
+      strategy = parsed.strategy;
+      tags = parsed.tags.map(normalizeTag).filter((t) => t && t.length >= 2);
+      return { strategy, tags };
+    }
+
+    // 旧格式: 纯数组
+    if (Array.isArray(parsed)) {
+      tags = parsed.map(normalizeTag).filter((t) => t && t.length >= 2);
+      return { strategy, tags };
+    }
+
+    // 旧格式: { tags: [...] }
+    if (Array.isArray(parsed.tags)) {
+      tags = parsed.tags.map(normalizeTag).filter((t) => t && t.length >= 2);
+      return { strategy, tags };
+    }
+  } catch {
+    // JSON 解析失败，尝试从文本中提取
+    // 先尝试提取 strategy（可能在 JSON 块之前）
+    const strategyMatch = content.match(/"strategy"\s*:\s*"([^"]+)"/);
+    if (strategyMatch) {
+      strategy = strategyMatch[1];
+    }
+  }
+
+  // 降级：从文本行中提取 tags
+  const lines = content.split(/[\n,]+/);
+  for (const line of lines) {
+    const cleaned = normalizeTag(line.replace(/^[-\d.\s]+/, ""));
+    if (cleaned && /^[a-z0-9_]+$/.test(cleaned) && cleaned.length >= 2) {
+      tags.push(cleaned);
+    }
+  }
+
+  return { strategy, tags };
+}
+
 // ====== Prompt 组装 ======
 
-export function buildDiscoverPrompt(country, count, history, userPrompt) {
+export function buildDiscoverPrompt(
+  country,
+  count,
+  history,
+  userPrompt,
+  pastStrategies,
+) {
   const lang = getLang(country);
   const langNames = {
     cs: "Czech",
@@ -98,11 +155,15 @@ export function buildDiscoverPrompt(country, count, history, userPrompt) {
   };
   const langName = langNames[lang] || lang;
 
-  // 正样本：该国高分 tag（只给 LLM 看效果，不给模板）
+  // 正样本：该国高分 tag（带分数，让 LLM 知道哪些方向最有效）
   const productive = history.productive || [];
   const productiveHint =
     productive.length > 0
-      ? `\nTags that already worked well for ${country}: ${productive.map((t) => t.tag).join(", ")}. These are examples of what works — explore DIFFERENT directions, not variations of these.`
+      ? `\nTags that already worked well for ${country} (with scores): ${productive
+          .map((t) => `${t.tag}(score:${Math.round(t.score)})`)
+          .join(
+            ", ",
+          )}. These are examples of what works — explore DIFFERENT directions, not variations of these.`
       : "";
 
   // 负样本：该国 dead tag
@@ -132,6 +193,23 @@ export function buildDiscoverPrompt(country, count, history, userPrompt) {
     ? `\nAdditional focus: ${userPrompt}. Generate tags specifically for this niche.`
     : "";
 
+  // 历史策略反馈：让 LLM 知道之前的策略效果如何
+  let strategyReview = "";
+  if (pastStrategies && pastStrategies.length > 0) {
+    const strategyLines = pastStrategies
+      .map(
+        (s) =>
+          `Round ${s.round}: strategy="${s.strategy?.substring(0, 200)}", added ${s.tags_added} tags, avg_score=${(s.avg_score || 0).toFixed(1)}, productive=${s.productive_count || 0}, dead=${s.dead_count || 0}`,
+      )
+      .join("\n");
+    strategyReview = `
+=== Previous discovery rounds (learn from these results) ===
+${strategyLines}
+
+Based on the above, which strategies produced high-scoring tags? Which failed?
+Use this analysis to decide your strategy for this round.`;
+  }
+
   return `You are discovering TikTok hashtags used by people who sell things in ${country}.
 
 Your goal: Find hashtags that real sellers in ${country} actually use — any kind of tag they might use. Think broadly:
@@ -139,16 +217,32 @@ Your goal: Find hashtags that real sellers in ${country} actually use — any ki
 - What they sell (shoes, clothes, jewelry, food, pets, furniture...)
 - How they sell (online, handmade, second-hand, local pickup...)
 - Product-specific tags (sneakers, dresses, cakes, necklaces...)
+- Niche categories: beauty, fitness, pets, plants, books, toys, music, art, photography, gardening, DIY, automotive, real estate...
 
 All tags must be in ${langName} language (or widely used in ${country}).
-Generate ${count} tags that are ALL DIFFERENT from each other and from any existing tags.
+Generate ${count} tags that are ALL DIFFERENT from each other and from any existing tags.${productiveHint}${deadHint}${errorHint}${existingHint}${userHint}${strategyReview}
+
+## IMPORTANT: Think first, then generate
+
+Before generating tags, analyze:
+1. Which tag directions scored highest? What makes them work?
+2. Which directions completely failed? Why?
+3. What seller niches are NOT yet covered? (e.g., if we have "shop" but no "bakery", "petstore", "bookshop"...)
+4. What specific direction will YOU explore this round? Be concrete.
+
+## Output format
+
+Return ONLY a JSON object with two fields:
+{
+  "strategy": "Your analysis and plan for this round (2-4 sentences). Explain what direction you're exploring and why.",
+  "tags": ["tag1", "tag2", "tag3", "tag4"]
+}
 
 Rules:
 - Each tag should explore a DIFFERENT angle — don't just swap country suffixes
 - Prefer specific and niche tags over generic ones (e.g., "vendozapatos" beats "vender")
-- Do NOT generate tags that already exist${productiveHint}${deadHint}${errorHint}${existingHint}${userHint}
-
-Return ONLY a JSON array of tag strings, nothing else. Example: ["ventas","tiendaonline","vender"]`;
+- Do NOT generate tags that already exist
+- Your strategy should be different from previous rounds — if a direction worked, go deeper; if it failed, try something new`;
 }
 
 // ====== discover: 单国家标签发现 ======
@@ -171,15 +265,32 @@ export async function discoverTagsForCountry(
   const allExisting = allTags.map((t) => t.tag);
   const history = { productive, dead, allExisting };
 
-  // 组装 prompt 并调用 LLM
-  const prompt = buildDiscoverPrompt(country, count, history, userPrompt);
+  // 读取历史策略效果（自学习）
+  const { getDiscoverRound, getDiscoverEffectiveness } =
+    await import("./db-tags.js");
+  const round = getDiscoverRound(country);
+  const pastStrategies = getDiscoverEffectiveness(country);
+
+  // 组装 prompt 并调用 LLM（两阶段：先思考策略，再生成 tag）
+  const prompt = buildDiscoverPrompt(
+    country,
+    count,
+    history,
+    userPrompt,
+    pastStrategies,
+  );
   console.error(
-    `[tag-service] LLM discover ${country} (lang=${getLang(country)}, count=${count})`,
+    `[tag-service] LLM discover ${country} round=${round} (lang=${getLang(country)}, count=${count})`,
   );
   const content = await callLLM(prompt);
-  const tags = parseTagsFromResponse(content);
+
+  // 解析响应：提取 strategy 和 tags
+  const { strategy, tags } = parseDiscoverResponse(content);
   const unique = [...new Set(tags)].slice(0, count);
 
+  console.error(
+    `[tag-service] LLM strategy: ${strategy?.substring(0, 150) || "(none)"}`,
+  );
   console.error(
     `[tag-service] LLM generated ${unique.length} tags: ${unique.join(", ")}`,
   );
@@ -191,8 +302,28 @@ export async function discoverTagsForCountry(
     if (result.inserted) inserted.push(result.tag);
   }
 
+  // 记录 discover log（策略 + 结果）
+  const { insertDiscoverLog } = await import("./db-tags.js");
+  insertDiscoverLog({
+    country,
+    round,
+    strategy,
+    tagsGenerated: unique,
+    tagsAdded: inserted.length,
+    productiveSample: productive.slice(0, 20).map((t) => t.tag),
+    deadSample: dead.slice(0, 20).map((t) => t.tag),
+    avgProductiveScore:
+      productive.length > 0
+        ? productive.reduce((s, t) => s + t.score, 0) / productive.length
+        : 0,
+    avgDeadScore:
+      dead.length > 0 ? dead.reduce((s, t) => s + t.score, 0) / dead.length : 0,
+  });
+
   return {
     country,
+    round,
+    strategy,
     added: inserted.length,
     tags: inserted,
     total: unique.length,