Expand chat prompts with GLM before submitting to H3

Adds a prompt-optimization step ahead of H3 generation: a viewer's short chat line ("一只猫跳舞") gets expanded into a detailed shot-by-shot English prompt (camera work, lighting, action, setting) via one GLM Coding Plan call (glm-5.3-flash, /api/coding/ endpoint — the coding-plan-specific path, not the pay-per-token one) before being handed to H3. The expanded prompt becomes the clip's fullPrompt; the original short text stays as the display prompt. Persisted on the queue entry so a restart mid-generation doesn't re-call GLM. Best-effort throughout — no API key, a timeout, or any GLM error just falls back to using the original text unchanged, never blocking generation. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0144MiJ2cLjMd9jWaeEpVfUJ

Expand chat prompts with GLM before submitting to H3
Adds a prompt-optimization step ahead of H3 generation: a viewer's short chat line ("一只猫跳舞") gets expanded into a detailed shot-by-shot English prompt (camera work, lighting, action, setting) via one GLM Coding Plan call (glm-5.3-flash, /api/coding/ endpoint — the coding-plan-specific path, not the pay-per-token one) before being handed to H3. The expanded prompt becomes the clip's fullPrompt; the original short text stays as the display prompt. Persisted on the queue entry so a restart mid-generation doesn't re-call GLM. Best-effort throughout — no API key, a timeout, or any GLM error just falls back to using the original text unchanged, never blocking generation. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0144MiJ2cLjMd9jWaeEpVfUJ
yibo
1 parent 1dc70844
Showing 1 changed file with 67 additions and 2 deletions
server.js
@@ -19,6 +19,15 @@ const H3_CONSOLE_BASE = process.env.H3_CONSOLE_BASE || &quot;&quot;;
 const H3_CONSOLE_TOKEN = process.env.H3_CONSOLE_TOKEN || "";
 const H3_CONSOLE_URL = H3_CONSOLE_BASE ? new URL(H3_CONSOLE_BASE) : null;
+// GLM Coding Plan — used only to expand a viewer's short chat prompt into a
+// detailed shot-by-shot prompt before handing it to H3 (see optimizePrompt).
+// Must hit the /api/coding/ path specifically — the plain /api/paas/ path
+// bills against a different (pay-per-token) quota than the coding plan.
+const GLM_API_KEY = process.env.GLM_API_KEY || "";
+const GLM_HOST = "open.bigmodel.cn";
+const GLM_PATH = "/api/coding/paas/v4/chat/completions";
+const GLM_MODEL = "glm-5.3-flash";
+
 const MIME = { ".html": "text/html", ".js": "text/javascript", ".css": "text/css", ".json": "application/json" };
 const { syncLibrary } = require("./ingest.js");
@@ -91,6 +100,58 @@ function h3ConsoleFetch(pathAndQuery, options) {
   });
 }
+function glmChat(messages) {
+  return new Promise((resolve, reject) => {
+    const body = JSON.stringify({ model: GLM_MODEL, messages, temperature: 0.7 });
+    const req = https.request(
+      {
+        hostname: GLM_HOST,
+        path: GLM_PATH,
+        method: "POST",
+        headers: { Authorization: `Bearer ${GLM_API_KEY}`, "Content-Type": "application/json", "Content-Length": Buffer.byteLength(body) },
+      },
+      (res) => {
+        let data = "";
+        res.on("data", (c) => (data += c));
+        res.on("end", () => {
+          try {
+            resolve(JSON.parse(data));
+          } catch (err) {
+            reject(err);
+          }
+        });
+      }
+    );
+    req.on("error", reject);
+    req.setTimeout(15000, () => req.destroy(new Error("GLM request timed out")));
+    req.write(body);
+    req.end();
+  });
+}
+
+const PROMPT_OPTIMIZE_SYSTEM = `You are a prompt engineer for a text-to-video model. Expand the user's short
+request (which may be in Chinese or English) into a detailed shot-by-shot generation prompt written in
+English: camera angle/movement, lighting and mood, the subject's specific actions, and the setting.
+Output only the expanded English prompt itself — no preamble, no quotes, no "Prompt:" prefix — under
+150 words.`;
+
+// Best-effort — a viewer's short chat line ("一只猫跳舞") makes for a weak
+// H3 prompt on its own. Never let a GLM outage or slow response block
+// generation: on any failure this just falls back to the original text.
+async function optimizePrompt(userPrompt) {
+  if (!GLM_API_KEY) return userPrompt;
+  try {
+    const res = await glmChat([
+      { role: "system", content: PROMPT_OPTIMIZE_SYSTEM },
+      { role: "user", content: userPrompt },
+    ]);
+    const text = res.choices && res.choices[0] && res.choices[0].message && res.choices[0].message.content;
+    return text ? text.trim() : userPrompt;
+  } catch {
+    return userPrompt;
+  }
+}
+
 // ---- Live channel: shared virtual playhead + prompt queue ----
 // Persisted to live-state.json (see saveLiveState/loadLiveState) so a
 // restart resumes the schedule instead of starting cold; ensureSchedule's
@@ -277,9 +338,13 @@ function isRateLimited(ip) {
 async function generateForPrompt(entry) {
   try {
     if (!entry.h3JobId) {
+      if (!entry.expandedPrompt) {
+        entry.expandedPrompt = await optimizePrompt(entry.prompt);
+        saveLiveState();
+      }
       const submitted = await h3ConsoleFetch("/v1/videos", {
         method: "POST",
-        body: { prompt: entry.prompt, ...H3_GEN_PARAMS },
+        body: { prompt: entry.expandedPrompt, ...H3_GEN_PARAMS },
       });
       if (!submitted.id) throw new Error(submitted.detail || "submit failed");
       entry.h3JobId = submitted.id;
@@ -295,7 +360,7 @@ async function generateForPrompt(entry) {
           id: job.id,
           source: "h3console",
           prompt: entry.prompt,
-          fullPrompt: entry.prompt,
+          fullPrompt: entry.expandedPrompt || entry.prompt,
           width: job.req.width,
           height: job.req.height,
           duration: Math.round((job.req.frames / FPS) * 10) / 10,