@modusensus/dsh-mneme 0.6.8 → 0.6.10

package/src/dream/tag-extractor.js

@@ -1,156 +1,156 @@
-// v0.6.2 auto-tag: a lightweight LLM pass that runs after an autoDream
-// consolidation and extracts 1-3 tags per retained memory. Opt-in via
-// config.autoTagEnabled (default false); autoTagMaxPerRun (default 10) bounds
-// how many memories are tagged per run.
-//
-// Design notes:
-//   - ONE batched LLM call per run (the consolidation route is reused). Each
-//     memory is a prompt line; the model replies with a JSON array of
-//     {"id","tags"} entries — the per-memory contract is `{"tags":[...]}`.
-//   - Fail-safe everywhere: a missing route / aborted stream / unparseable
-//     JSON / unknown id / illegal tag are all skipped, never thrown. Tagging
-//     must never degrade the consolidation run it rides on.
-//   - The actual writes go through service.applyMemoryTags inside a
-//     service.transaction, so the mirror re-renders exactly once per pass and
-//     the write hooks fire once (never per memory).
-import { sanitizeTags, MAX_TAG_LENGTH } from "../parser/tag.js";
-
-const TAG_PROMPT = `你是记忆库标签助手。下面是保留的记忆条目（id、标题、内容）。
-对每条记忆提取 1-3 个中文或英文标签，用于检索分类。
-标签规则：
-- 只允许字符：字母、数字、下划线、中文、连字符（如：linux、考研、deepseek-r1）
-- 标签长度 ≤ ${MAX_TAG_LENGTH} 字符
-- 宁缺毋滥：提取最核心的 1-3 个，不要凑数
-- 不要输出内容里没有依据的标签
-只输出一个 JSON 数组，每项形如 { "id": "<记忆id>", "tags": ["标签1", "标签2"] }。
-不要输出其他文字。`;
-
-/** Same stream consumption contract as dream.js. Returns accumulated text or
- *  undefined when the stream aborted/errored. */
-async function streamText(ctx, options) {
-  if (!ctx?.llm?.stream) return undefined;
-  let text = "";
-  for await (const chunk of ctx.llm.stream(options)) {
-    if (chunk.type === "text-delta" && typeof chunk.text === "string") text += chunk.text;
-    if (chunk.type === "finish" && (chunk.reason?.kind === "error" || chunk.reason?.kind === "aborted")) {
-      return undefined;
-    }
-  }
-  return text;
-}
-
-/** Pull the outermost JSON array out of a model reply (same tolerant contract
- *  as sleep.js parseJsonArray): find the first `[` … last `]` and parse. */
-function parseJsonArray(text) {
-  if (typeof text !== "string") return undefined;
-  const start = text.indexOf("[");
-  const end = text.lastIndexOf("]");
-  if (start === -1 || end <= start) return undefined;
-  try {
-    const parsed = JSON.parse(text.slice(start, end + 1));
-    return Array.isArray(parsed) ? parsed : undefined;
-  } catch {
-    return undefined;
-  }
-}
-
-/** Resolve the LLM route for the tag pass: the caller-provided consolidation
- *  route first, then dreamProvider/dreamModel. Falls through to undefined. */
-function resolveTagRoute(route, config, logger) {
-  if (route?.provider && route?.model) return route;
-  if (config?.dreamProvider && config?.dreamModel) return { provider: config.dreamProvider, model: config.dreamModel };
-  logger?.warn?.("dsh-mneme auto-tag: no llm route available");
-  return undefined;
-}
-
-/**
- * Run the auto-tag pass over retained memories.
- * @param {object} opts
- * @param {object} opts.ctx        — { llm, logger }
- * @param {object} opts.service    — service handle (all/transaction/applyMemoryTags)
- * @param {object} opts.config     — plugin config (autoTagMaxPerRun, dreamProvider…)
- * @param {object} [opts.route]    — already-resolved consolidation route
- * @returns {Promise<{ok: boolean, tagged: number, skipped: number, failed: number, skippedBy: boolean}>}
- */
-export async function runAutoTag({ ctx, service, config, route }) {
-  const logger = ctx?.logger;
-  const maxPerRun = Number.isInteger(config?.autoTagMaxPerRun) && config.autoTagMaxPerRun > 0
-    ? config.autoTagMaxPerRun
-    : 10;
-  // Retained = post-consolidation active memories, newest first, capped.
-  // Skip memories that already carry live tags — otherwise every run re-picks
-  // the same "newest" batch (tagging doesn't bump updated_at) and older
-  // memories starve; manual tags are also left untouched this way.
-  const memories = service.all()
-    .filter((m) => !m.forgotten && !m.archived && !m.session_disposed_at && m.type !== "summary")
-    .filter((m) => (service.getMemoryTags?.(m.id) ?? []).length === 0)
-    .sort((a, b) => {
-      const ta = String(a.updated_at ?? "");
-      const tb = String(b.updated_at ?? "");
-      if (ta < tb) return 1;
-      if (ta > tb) return -1;
-      return a.id < b.id ? -1 : a.id > b.id ? 1 : 0;
-    })
-    .slice(0, Math.max(1, maxPerRun));
-  if (!memories.length) return { ok: true, tagged: 0, skipped: 0, failed: 0, skippedBy: "empty" };
-  const tagRoute = resolveTagRoute(route, config, logger);
-  if (!tagRoute) return { ok: false, tagged: 0, skipped: memories.length, failed: 0, skippedBy: "no-route" };
-
-  const listText = memories
-    .map((m) => `id=${m.id} | title=${m.title} | content=${m.content}`)
-    .join("\n");
-  const text = await streamText(ctx, {
-    provider: tagRoute.provider,
-    model: tagRoute.model,
-    purpose: "compaction",
-    maxTokens: Math.min(2048, config?.dreamMaxTokens ?? 2048),
-    messages: [
-      { role: "system", content: [{ type: "text", text: TAG_PROMPT }] },
-      { role: "user", content: [{ type: "text", text: listText }] }
-    ]
-  });
-  if (text === undefined) return { ok: false, tagged: 0, skipped: memories.length, failed: 0, skippedBy: "llm-failed" };
-
-  const entries = parseJsonArray(text);
-  if (!entries) {
-    logger?.warn?.("dsh-mneme auto-tag: no json array in llm output");
-    return { ok: false, tagged: 0, skipped: memories.length, failed: 0, skippedBy: "bad-json" };
-  }
-
-  // Validate ids against the candidate set (unknown ids are ignored, never
-  // written — a stray id could otherwise tag an unrelated memory).
-  const candidateIds = new Set(memories.map((m) => m.id));
-  const toWrite = [];
-  let skipped = 0;
-  let failed = 0;
-  const seenIds = new Set();
-  for (const entry of entries) {
-    if (!entry || typeof entry !== "object") { failed++; continue; }
-    const id = typeof entry.id === "string" ? entry.id : undefined;
-    if (!id || !candidateIds.has(id)) { failed++; continue; }
-    if (seenIds.has(id)) continue; // first entry per id wins
-    seenIds.add(id);
-    // Reuse the shared sanitizer: strip, ≤20 chars, drop illegal, dedupe.
-    const tags = sanitizeTags(entry.tags);
-    if (!tags.length) { skipped++; continue; }
-    toWrite.push({ id, tags });
-  }
-
-  // Write the whole batch under one transaction: mirror re-renders once.
-  let tagged = 0;
-  try {
-    service.transaction(() => {
-      for (const { id, tags } of toWrite) {
-        // Merge with existing live tags (never clobber manual tags that landed
-        // between selection and write).
-        const existing = service.getMemoryTags?.(id) ?? [];
-        service.applyMemoryTags(id, [...new Set([...existing, ...tags])]);
-        tagged++;
-      }
-    });
-  } catch (error) {
-    logger?.warn?.(`dsh-mneme auto-tag: write failed: ${String(error)}`);
-    return { ok: false, tagged, skipped, failed, skippedBy: "write-failed" };
-  }
-  return { ok: true, tagged, skipped, failed, skippedBy: false };
-}
+// v0.6.2 auto-tag: a lightweight LLM pass that runs after an autoDream
+// consolidation and extracts 1-3 tags per retained memory. Opt-in via
+// config.autoTagEnabled (default false); autoTagMaxPerRun (default 10) bounds
+// how many memories are tagged per run.
+//
+// Design notes:
+//   - ONE batched LLM call per run (the consolidation route is reused). Each
+//     memory is a prompt line; the model replies with a JSON array of
+//     {"id","tags"} entries — the per-memory contract is `{"tags":[...]}`.
+//   - Fail-safe everywhere: a missing route / aborted stream / unparseable
+//     JSON / unknown id / illegal tag are all skipped, never thrown. Tagging
+//     must never degrade the consolidation run it rides on.
+//   - The actual writes go through service.applyMemoryTags inside a
+//     service.transaction, so the mirror re-renders exactly once per pass and
+//     the write hooks fire once (never per memory).
+import { sanitizeTags, MAX_TAG_LENGTH } from "../parser/tag.js";
+
+const TAG_PROMPT = `你是记忆库标签助手。下面是保留的记忆条目（id、标题、内容）。
+对每条记忆提取 1-3 个中文或英文标签，用于检索分类。
+标签规则：
+- 只允许字符：字母、数字、下划线、中文、连字符（如：linux、考研、deepseek-r1）
+- 标签长度 ≤ ${MAX_TAG_LENGTH} 字符
+- 宁缺毋滥：提取最核心的 1-3 个，不要凑数
+- 不要输出内容里没有依据的标签
+只输出一个 JSON 数组，每项形如 { "id": "<记忆id>", "tags": ["标签1", "标签2"] }。
+不要输出其他文字。`;
+
+/** Same stream consumption contract as dream.js. Returns accumulated text or
+ *  undefined when the stream aborted/errored. */
+async function streamText(ctx, options) {
+  if (!ctx?.llm?.stream) return undefined;
+  let text = "";
+  for await (const chunk of ctx.llm.stream(options)) {
+    if (chunk.type === "text-delta" && typeof chunk.text === "string") text += chunk.text;
+    if (chunk.type === "finish" && (chunk.reason?.kind === "error" || chunk.reason?.kind === "aborted")) {
+      return undefined;
+    }
+  }
+  return text;
+}
+
+/** Pull the outermost JSON array out of a model reply (same tolerant contract
+ *  as sleep.js parseJsonArray): find the first `[` … last `]` and parse. */
+function parseJsonArray(text) {
+  if (typeof text !== "string") return undefined;
+  const start = text.indexOf("[");
+  const end = text.lastIndexOf("]");
+  if (start === -1 || end <= start) return undefined;
+  try {
+    const parsed = JSON.parse(text.slice(start, end + 1));
+    return Array.isArray(parsed) ? parsed : undefined;
+  } catch {
+    return undefined;
+  }
+}
+
+/** Resolve the LLM route for the tag pass: the caller-provided consolidation
+ *  route first, then dreamProvider/dreamModel. Falls through to undefined. */
+function resolveTagRoute(route, config, logger) {
+  if (route?.provider && route?.model) return route;
+  if (config?.dreamProvider && config?.dreamModel) return { provider: config.dreamProvider, model: config.dreamModel };
+  logger?.warn?.("dsh-mneme auto-tag: no llm route available");
+  return undefined;
+}
+
+/**
+ * Run the auto-tag pass over retained memories.
+ * @param {object} opts
+ * @param {object} opts.ctx        — { llm, logger }
+ * @param {object} opts.service    — service handle (all/transaction/applyMemoryTags)
+ * @param {object} opts.config     — plugin config (autoTagMaxPerRun, dreamProvider…)
+ * @param {object} [opts.route]    — already-resolved consolidation route
+ * @returns {Promise<{ok: boolean, tagged: number, skipped: number, failed: number, skippedBy: boolean}>}
+ */
+export async function runAutoTag({ ctx, service, config, route }) {
+  const logger = ctx?.logger;
+  const maxPerRun = Number.isInteger(config?.autoTagMaxPerRun) && config.autoTagMaxPerRun > 0
+    ? config.autoTagMaxPerRun
+    : 10;
+  // Retained = post-consolidation active memories, newest first, capped.
+  // Skip memories that already carry live tags — otherwise every run re-picks
+  // the same "newest" batch (tagging doesn't bump updated_at) and older
+  // memories starve; manual tags are also left untouched this way.
+  const memories = service.all()
+    .filter((m) => !m.forgotten && !m.archived && !m.session_disposed_at && m.type !== "summary")
+    .filter((m) => (service.getMemoryTags?.(m.id) ?? []).length === 0)
+    .sort((a, b) => {
+      const ta = String(a.updated_at ?? "");
+      const tb = String(b.updated_at ?? "");
+      if (ta < tb) return 1;
+      if (ta > tb) return -1;
+      return a.id < b.id ? -1 : a.id > b.id ? 1 : 0;
+    })
+    .slice(0, Math.max(1, maxPerRun));
+  if (!memories.length) return { ok: true, tagged: 0, skipped: 0, failed: 0, skippedBy: "empty" };
+  const tagRoute = resolveTagRoute(route, config, logger);
+  if (!tagRoute) return { ok: false, tagged: 0, skipped: memories.length, failed: 0, skippedBy: "no-route" };
+
+  const listText = memories
+    .map((m) => `id=${m.id} | title=${m.title} | content=${m.content}`)
+    .join("\n");
+  const text = await streamText(ctx, {
+    provider: tagRoute.provider,
+    model: tagRoute.model,
+    purpose: "compaction",
+    maxTokens: Math.min(2048, config?.dreamMaxTokens ?? 2048),
+    messages: [
+      { role: "system", content: [{ type: "text", text: TAG_PROMPT }] },
+      { role: "user", content: [{ type: "text", text: listText }] }
+    ]
+  });
+  if (text === undefined) return { ok: false, tagged: 0, skipped: memories.length, failed: 0, skippedBy: "llm-failed" };
+
+  const entries = parseJsonArray(text);
+  if (!entries) {
+    logger?.warn?.("dsh-mneme auto-tag: no json array in llm output");
+    return { ok: false, tagged: 0, skipped: memories.length, failed: 0, skippedBy: "bad-json" };
+  }
+
+  // Validate ids against the candidate set (unknown ids are ignored, never
+  // written — a stray id could otherwise tag an unrelated memory).
+  const candidateIds = new Set(memories.map((m) => m.id));
+  const toWrite = [];
+  let skipped = 0;
+  let failed = 0;
+  const seenIds = new Set();
+  for (const entry of entries) {
+    if (!entry || typeof entry !== "object") { failed++; continue; }
+    const id = typeof entry.id === "string" ? entry.id : undefined;
+    if (!id || !candidateIds.has(id)) { failed++; continue; }
+    if (seenIds.has(id)) continue; // first entry per id wins
+    seenIds.add(id);
+    // Reuse the shared sanitizer: strip, ≤20 chars, drop illegal, dedupe.
+    const tags = sanitizeTags(entry.tags);
+    if (!tags.length) { skipped++; continue; }
+    toWrite.push({ id, tags });
+  }
+
+  // Write the whole batch under one transaction: mirror re-renders once.
+  let tagged = 0;
+  try {
+    service.transaction(() => {
+      for (const { id, tags } of toWrite) {
+        // Merge with existing live tags (never clobber manual tags that landed
+        // between selection and write).
+        const existing = service.getMemoryTags?.(id) ?? [];
+        service.applyMemoryTags(id, [...new Set([...existing, ...tags])]);
+        tagged++;
+      }
+    });
+  } catch (error) {
+    logger?.warn?.(`dsh-mneme auto-tag: write failed: ${String(error)}`);
+    return { ok: false, tagged, skipped, failed, skippedBy: "write-failed" };
+  }
+  return { ok: true, tagged, skipped, failed, skippedBy: false };
+}