@cup319/mmpl 3.3.4 → 3.3.6

package/README.md

@@ -2,7 +2,7 @@
 
 MMPL (Memory Palace) — DuckDB 向量记忆 MCP Server（单二进制，Rust 实现）。
 
-当前版本 **3.3.4**：版本链（`memory_manage action=history`）、长记忆切块嵌入、管道幂等去重、LongMemEval 检索评测。
+当前版本 **3.3.6**：版本链（`memory_manage action=history`）、长记忆切块嵌入、管道幂等去重、LongMemEval 检索评测。
 
 ```bash
 npm install -g @cup319/mmpl

package/install.js

@@ -12,7 +12,7 @@ const { get } = require("https");
 const path = require("path");
 const { execSync } = require("child_process");
 
-const VERSION = "3.3.4";
+const VERSION = "3.3.6";
 const IS_WIN = process.platform === "win32";
 const PLATFORM = IS_WIN ? "windows" : "linux";
 const ARCH = "x86_64";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cup319/mmpl",
-  "version": "3.3.4",
+  "version": "3.3.6",
   "description": "MMPL (Memory Palace) — 单二进制 DuckDB 记忆 MCP Server (9 工具 + ONNX 语义向量 + 冲突双通道门控 + 场景构建 + 四层钻取) — Linux + Windows 双平台",
   "license": "MIT",
   "publishConfig": {
@@ -8,14 +8,16 @@
     "registry": "https://registry.npmjs.org/"
   },
   "bin": {
-    "mmpl": "run.js"
+    "mmpl": "run.js",
+    "mmpl-setup": "setup.js"
   },
   "files": [
     "install.js",
     "run.js",
+    "setup.js",
     "README.md",
-    "mmpl-v3.3.4-linux-x86_64.tar.gz",
-    "mmpl-v3.3.4-windows-x86_64.zip"
+    "mmpl-v3.3.6-linux-x86_64.tar.gz",
+    "mmpl-v3.3.6-windows-x86_64.zip"
   ],
   "scripts": {
     "postinstall": "node install.js"

package/run.js

@@ -38,6 +38,23 @@ if (process.platform === "win32" && !process.env.PORT) {
   probe.on("timeout", () => probe.destroy());
 }
 
+// CNB 云开发环境: 自动绑 0.0.0.0（CNB 端口代理拨容器 IP，127.0.0.1 会 connection refused）
+// 同时提醒安全: 0.0.0.0 + 无 token = 任何拿到 URL 的人都能访问管理后台
+const isCNB = !!process.env.CNB_REPO_NAME;
+if (isCNB && !process.env.HOST) {
+  process.env.HOST = "0.0.0.0";
+  console.log("[mmpl] 检测到 CNB 云开发环境 → HOST=0.0.0.0（供端口代理访问）");
+  if (!process.env.MMPL_AUTH_TOKEN) {
+    const tok = require("crypto").randomBytes(16).toString("hex");
+    console.log("");
+    console.log("[mmpl] ⚠ 安全提示: 已监听 0.0.0.0，建议设置访问令牌。本次已自动生成:");
+    console.log(`[mmpl]    MMPL_AUTH_TOKEN=${tok}`);
+    console.log(`[mmpl]    管理后台: https://<你的CNB端口域名>/admin?token=${tok}`);
+    console.log("[mmpl]    (将该 token 写入环境变量重启后固定; 不设置则后台无认证暴露)");
+    console.log("");
+  }
+}
+
 const child = spawn(BIN, process.argv.slice(2), {
   stdio: "inherit",
   env: process.env,

package/setup.js

@@ -0,0 +1,244 @@
+#!/usr/bin/env node
+/**
+ * mmpl-setup — 语义向量模型预下载器（多线程分块 + 进度条 + 镜像回退）
+ *
+ * 用法:
+ *   mmpl-setup                交互式（或 --yes 直接确认）
+ *   mmpl-setup --model small  指定档位 small|base|m3|qwen3（默认 base）
+ *   mmpl-setup --dir D:\mmpl  指定 DATA_DIR（默认读环境变量，再退 ./data）
+ *
+ * 下载内容: DATA_DIR/models/<model>.onnx + tokenizer.json
+ * 加速: HTTP Range 分块（hf-mirror 支持，302 到 CDN），8 并发；
+ *       镜像顺序 hf-mirror.com → huggingface.co（直连可用时）
+ */
+const fs = require("fs");
+const path = require("path");
+const https = require("https");
+const { execSync } = require("child_process");
+
+// ---------- 参数 ----------
+const argv = process.argv.slice(2);
+const yes = argv.includes("--yes") || argv.includes("-y");
+const modelKey = (argv.find((a, i) => argv[i - 1] === "--model") || "base").toLowerCase();
+const dirArg = argv.find((a, i) => argv[i - 1] === "--dir");
+
+const MODELS = {
+  small: { dim: 512, repo: "Xenova/bge-small-zh-v1.5", file: "bge-small-zh-v1.5-q.onnx", minSize: 25 * 1024 * 1024 },
+  base:  { dim: 768, repo: "Xenova/bge-base-zh-v1.5",  file: "bge-base-zh-v1.5-q.onnx",  minSize: 90 * 1024 * 1024 },
+  m3:    { dim: 1024, repo: "Xenova/bge-m3",            file: "bge-m3-q.onnx",            minSize: 200 * 1024 * 1024 },
+  qwen3: { dim: 1024, repo: "onnx-community/Qwen3-Embedding-0.6B-ONNX", file: "qwen3-embedding-0.6b-q.onnx", minSize: 300 * 1024 * 1024 },
+};
+const M = MODELS[modelKey];
+if (!M) { console.error(`[mmpl-setup] 未知档位: ${modelKey}（可选 small|base|m3|qwen3）`); process.exit(1); }
+
+const DATA_DIR = dirArg || process.env.DATA_DIR || path.join(process.cwd(), "data");
+const MODELS_DIR = path.join(DATA_DIR, "models");
+// 下载源顺序: HF_ENDPOINT 自定义 > 测速选快 > 默认国内镜像优先
+const ALL_MIRRORS = ["https://hf-mirror.com", "https://huggingface.co"];
+const CUSTOM = (process.env.HF_ENDPOINT || "").trim().replace(/\/+$/, "");
+const MIRRORS = CUSTOM ? [CUSTOM, ...ALL_MIRRORS] : ALL_MIRRORS;
+const CHUNK_WORKERS = 8;
+let ranked = []; // 测速后的有序镜像（main() 里填充）
+
+if (process.platform === "win32") {
+  try { execSync("chcp 65001", { stdio: "ignore", windowsHide: true }); } catch (_) {}
+}
+
+// ---------- 工具 ----------
+const mb = (n) => (n / 1024 / 1024).toFixed(1);
+function get(url, redirects = 5) {
+  return new Promise((resolve, reject) => {
+    https.get(url, { headers: { "User-Agent": "mmpl-setup" } }, (res) => {
+      if (res.statusCode >= 300 && res.statusCode < 400 && res.headers.location && redirects > 0) {
+        res.resume();
+        const next = new URL(res.headers.location, url).toString();
+        return resolve(get(next, redirects - 1));
+      }
+      if (res.statusCode !== 200 && res.statusCode !== 206) {
+        res.resume();
+        return reject(new Error(`HTTP ${res.statusCode} @ ${url.slice(0, 80)}`));
+      }
+      resolve(res);
+    }).on("error", reject);
+  });
+}
+async function headSize(url) {
+  return new Promise((resolve, reject) => {
+    https.request(url, { method: "HEAD", headers: { "User-Agent": "mmpl-setup" } }, (res) => {
+      res.resume();
+      if (res.statusCode >= 300 && res.statusCode < 400 && res.headers.location) {
+        return resolve(headSize(new URL(res.headers.location, url).toString()));
+      }
+      const len = Number(res.headers["content-length"] || 0);
+      len > 0 ? resolve(len) : reject(new Error("no content-length"));
+    }).on("error", reject).end();
+  });
+}
+function fmtProgress(done, total, speed) {
+  const pct = total ? Math.floor((done / total) * 100) : 0;
+  const bar = "#".repeat(Math.floor(pct / 3)).padEnd(33, ".");
+  return `${bar} ${pct}% ${mb(done)}/${mb(total)}MB ${speed || ""}`;
+}
+/// 跟随 302 拿最终 CDN URL（签名 URL 才能带 Range 复用）
+function resolveUrl(url, redirects = 5) {
+  return new Promise((resolve, reject) => {
+    https.request(url, { method: "GET", headers: { "User-Agent": "mmpl-setup", Range: "bytes=0-0" } }, (res) => {
+      res.resume();
+      if (res.statusCode >= 300 && res.statusCode < 400 && res.headers.location && redirects > 0) {
+        return resolve(resolveUrl(new URL(res.headers.location, url).toString(), redirects - 1));
+      }
+      resolve(url); // 200/206 → 已是最终地址
+    }).on("error", reject).end();
+  });
+}
+
+// ---------- 分块并行下载 ----------
+async function downloadFile(url, dest, label) {
+  const total = await headSize(url);
+  // 小文件或服务器不支持 Range 时直接单线程
+  const supportsRange = true; // hf-mirror 与 hf CDN 均支持；失败自动走单线程回退
+  if (!supportsRange || total < 4 * 1024 * 1024) return downloadSingle(url, dest, label);
+
+  const partSize = Math.ceil(total / CHUNK_WORKERS);
+  const parts = [];
+  for (let i = 0; i < CHUNK_WORKERS; i++) {
+    const start = i * partSize;
+    if (start >= total) break;
+    const end = Math.min(start + partSize - 1, total - 1);
+    parts.push({ i, start, end, path: `${dest}.part${i}` });
+  }
+  let done = parts.filter((p) => fs.existsSync(p.path)).reduce((s, p) => s + (fs.statSync(p.path).size), 0);
+  const t0 = Date.now();
+  const spinner = setInterval(() => {
+    done = parts.reduce((s, p) => s + (fs.existsSync(p.path) ? fs.statSync(p.path).size : 0), 0);
+    const secs = (Date.now() - t0) / 1000;
+    process.stdout.write(`\r  ${label} ${fmtProgress(done, total, `${mb(done / secs)}MB/s`)}`.padEnd(100));
+  }, 300);
+
+  await Promise.all(parts.map((p) =>
+    (async () => {
+      const have = fs.existsSync(p.path) ? fs.statSync(p.path).size : 0;
+      const want = p.end - p.start + 1;
+      if (have === want) return; // 断点续传
+      const finalUrl = await resolveUrl(url); // 每分块解析一次最终 CDN URL（签名时效不同）
+      for (let attempt = 0; attempt < 3; attempt++) {
+        try {
+          // 带 Range 的原生请求
+          return await new Promise((resolve, reject) => {
+            const req = https.request(new URL(finalUrl), { headers: { "User-Agent": "mmpl-setup", Range: `bytes=${p.start + have}-${p.end}` } }, (res) => {
+              if (res.statusCode !== 206) { res.resume(); return reject(new Error(`part ${p.i}: HTTP ${res.statusCode}（镜像不支持断点）`)); }
+              const out = fs.createWriteStream(p.path, { flags: have ? "a" : "w" });
+              res.pipe(out);
+              out.on("finish", resolve).on("error", reject);
+            });
+            req.on("error", reject).end();
+          });
+        } catch (e) { if (attempt === 2) throw e; await new Promise((r) => setTimeout(r, 1000 * (attempt + 1))); }
+      }
+    })()
+  ));
+  clearInterval(spinner);
+  process.stdout.write("\n");
+  // 合并分块
+  const out = fs.createWriteStream(dest);
+  for (const p of parts) {
+    await new Promise((resolve, reject) => {
+      fs.createReadStream(p.path).on("data", (c) => out.write(c)).on("end", resolve).on("error", reject);
+    });
+    fs.unlinkSync(p.path);
+  }
+  await new Promise((r) => out.end(r));
+  if (fs.statSync(dest).size !== total) throw new Error(`大小不符: ${fs.statSync(dest).size} != ${total}`);
+}
+
+async function downloadSingle(url, dest, label) {
+  const res = await get(url);
+  const total = Number(res.headers["content-length"] || 0);
+  let done = 0; const t0 = Date.now();
+  const spinner = setInterval(() => {
+    const secs = (Date.now() - t0) / 1000;
+    process.stdout.write(`\r  ${label} ${fmtProgress(done, total, `${mb(done / secs)}MB/s`)}`.padEnd(100));
+  }, 300);
+  await new Promise((resolve, reject) => {
+    const out = fs.createWriteStream(dest);
+    res.on("data", (c) => { done += c.length; out.write(c); });
+    res.on("end", () => out.end(resolve)).on("error", reject);
+    out.on("error", reject);
+  });
+  clearInterval(spinner);
+  process.stdout.write("\n");
+}
+
+/// 并行测速（拉 1MB 样本），返回按速度排序的镜像列表（国内外自动择优）
+async function rankMirrors() {
+  const results = await Promise.all(MIRRORS.map(async (host) => {
+    const t0 = Date.now();
+    try {
+      // 用各仓库都存在的小文件测：tokenizer.json（~100-700KB）
+      const res = await get(`${host}/${M.repo}/resolve/main/tokenizer.json`);
+      const chunks = [];
+      for await (const c of res) chunks.push(c);
+      const bytes = chunks.reduce((s, c) => s + c.length, 0);
+      const secs = (Date.now() - t0) / 1000;
+      return { host, ok: true, speed: bytes / 1024 / 1024 / secs };
+    } catch { return { host, ok: false, speed: 0 }; }
+  }));
+  const ok = results.filter((r) => r.ok).sort((a, b) => b.speed - a.speed);
+  if (ok.length > 1) {
+    console.log("  测速:");
+    for (const r of ok) console.log(`    ${r.host}  ${r.speed.toFixed(2)} MB/s`);
+  }
+  return ok.map((r) => r.host);
+}
+
+// ---------- 主流程 ----------
+async function tryMirrors(buildUrl, dest, label) {
+  let lastErr;
+  for (const mirror of ranked) {
+    const url = buildUrl(mirror);
+    try {
+      await downloadFile(url, dest, label);
+      return mirror;
+    } catch (e) { lastErr = e; console.error(`\n  [镜像失败] ${mirror}: ${e.message}`); }
+  }
+  throw lastErr;
+}
+
+async function main() {
+  const modelPath = path.join(MODELS_DIR, M.file);
+  const tokPath = path.join(MODELS_DIR, "tokenizer.json");
+
+  if (fs.existsSync(modelPath) && fs.statSync(modelPath).size > M.minSize / 2 && fs.existsSync(tokPath)) {
+    console.log(`[mmpl-setup] 模型已就绪: ${modelPath} (${mb(fs.statSync(modelPath).size)}MB)`);
+    console.log("[mmpl-setup] 无需下载，退出。");
+    return;
+  }
+
+  console.log("[mmpl-setup] MMPL 语义向量模型预下载");
+  console.log(`  档位: ${modelKey} (${M.dim} 维)`);
+  console.log(`  目标: ${modelPath}`);
+  console.log(`  体积: ~${Math.round(M.minSize / 1024 / 1024)}MB+，${CHUNK_WORKERS} 线程分块下载`);
+  console.log(`  源: ${CUSTOM ? `自定义 ${CUSTOM} + 回退` : "国内 hf-mirror.com / 海外 huggingface.co 自动择优"}`);
+  console.log("  跳过则首次启动时由 mmpl 单线程下载（较慢）");
+  console.log("");
+  if (!yes && process.stdin.isTTY) {
+    process.stdout.write("现在下载? [Y/n] ");
+    const answer = await new Promise((r) => process.stdin.once("data", (d) => r(d.toString().trim().toLowerCase())));
+    if (answer === "n" || answer === "no") { console.log("已跳过。需要时运行: mmpl-setup --yes"); return; }
+  }
+
+  fs.mkdirSync(MODELS_DIR, { recursive: true });
+  ranked = await rankMirrors();
+  if (!ranked.length) { console.error("[mmpl-setup] 所有镜像均不可达，请检查网络或设置 HF_ENDPOINT"); process.exit(1); }
+  console.log(`  使用源: ${ranked[0]}${ranked.length > 1 ? `（回退: ${ranked.slice(1).join(", ")}）` : ""}`);
+  const t0 = Date.now();
+  const m = await tryMirrors((host) => `${host}/${M.repo}/resolve/main/onnx/model_quantized.onnx`, modelPath, "模型");
+  console.log(`  模型完成（${m}，${mb(fs.statSync(modelPath).size)}MB，${((Date.now() - t0) / 1000).toFixed(1)}s）`);
+  await tryMirrors((host) => `${host}/${M.repo}/resolve/main/tokenizer.json`, tokPath, "分词器");
+  console.log(`  分词器完成（${mb(fs.statSync(tokPath).size)}MB）`);
+  console.log("");
+  console.log("[mmpl-setup] 完成！mmpl 启动将直接使用本地模型。");
+  console.log(`[mmpl-setup] 提示: 环境变量 MMPL_MODEL=${modelKey}（默认 base）`);
+}
+
+main().catch((e) => { console.error(`[mmpl-setup] 下载失败: ${e.message}`); console.error("[mmpl-setup] 可稍后重试（支持断点续传），或检查网络代理设置"); process.exit(1); });