Increase translation batch size from 3 to 8

Send eight semantic segments per request so a typical viewport finishes in one round-trip. Raise the background cap and max tokens to match.

Author: qydxyx

README.md

@@ -134,9 +134,9 @@ Current as of August 10, 2026, DeepSeek lists the following prices per 1 million
 
 DeepSeek says these prices may increase soon, so check the current pricing page before relying on this estimate. Its official [token usage guide](https://api-docs.deepseek.com/quick_start/token_usage/) estimates about 0.3 token per English character and about 0.6 token per Chinese character. Its [context caching guide](https://api-docs.deepseek.com/guides/kv_cache/) explains the automatic best-effort disk cache used for repeated prefixes.
 
-A measured 20-minute English talk contained **2,935 spoken English words** and 15,433 transcript characters. With YouTube Digest's current grouping, it became 128 semantic segments and 43 requests of three segments each. Repeated prompts and JSON brought the rendered input to about 108,528 English characters, or **about 32,600 input tokens** using DeepSeek's 0.3 token per English character heuristic. The translated Chinese JSON output is estimated at about 3,500 to 4,500 tokens using the 0.6 token per Chinese character heuristic, plus JSON and ID overhead.
+A measured 20-minute English talk contained **2,935 spoken English words** and 15,433 transcript characters. With YouTube Digest's current grouping, it became 128 semantic segments and 16 requests of eight segments each. Repeated prompts and JSON bring the rendered input to about 50,100 English characters, or **about 15,000 input tokens** using DeepSeek's 0.3 token per English character heuristic. The translated Chinese JSON output is estimated at about 3,500 to 4,500 tokens using the 0.6 token per Chinese character heuristic, plus JSON and ID overhead.
 
-If all input is billed as cache miss, input costs about $0.0046 and output costs about $0.0010 to $0.0013, for a total of about $0.0056 to $0.0059. When much of the repeated system prompt hits DeepSeek's automatic best-effort cache, a realistic lower end is about $0.002 to $0.003. A practical estimate for fully translating this talk is therefore **$0.002 to $0.006 USD, about ¥0.02 to ¥0.04**.
+If all input is billed as cache miss, input costs about $0.0021 and output costs about $0.0010 to $0.0013, for a total of about $0.0031 to $0.0034. When much of the repeated system prompt hits DeepSeek's automatic best-effort cache, a realistic lower end is about $0.001 to $0.002. A practical estimate for fully translating this talk is therefore **$0.001 to $0.004 USD, about ¥0.01 to ¥0.03**.
 
 Translation is lazy and progressive. Cached segments are reused, and only rows you request by scrolling into them incur calls. Retries, provider behavior, and pricing changes can increase the final cost.
 

README.zh-CN.md

@@ -134,9 +134,9 @@ DeepSeek 的额度与 Supadata 分开计算。DeepSeek 可能有自己的免费
 
 DeepSeek 说明这些价格可能很快上调，因此使用此估算前必须查看当前价格页面。官方 [token 用量指南](https://api-docs.deepseek.com/quick_start/token_usage/)估算每个英文字符约为 0.3 token，每个中文字符约为 0.6 token。[上下文缓存指南](https://api-docs.deepseek.com/guides/kv_cache/)说明了重复前缀使用的自动尽力而为磁盘缓存。
 
-一个实测的 20 分钟英文演讲包含 **2,935 个英文口语词**和 15,433 个字幕字符。按 YouTube Digest 当前的分组方式，它会变成 128 个语义分段，以每次 3 段的方式发出 43 次请求。算上重复 prompt 和 JSON 后，渲染后的输入约为 108,528 个英文字符，按官方每个英文字符 0.3 token 的经验值，即**约 32,600 个输入 token**。按每个中文字符 0.6 token 的经验值，再加上 JSON 和 ID 开销，中文 JSON 输出估计为 3,500 到 4,500 token。
+一个实测的 20 分钟英文演讲包含 **2,935 个英文口语词**和 15,433 个字幕字符。按 YouTube Digest 当前的分组方式，它会变成 128 个语义分段，以每次 8 段的方式发出 16 次请求。算上重复 prompt 和 JSON 后，渲染后的输入约为 50,100 个英文字符，按官方每个英文字符 0.3 token 的经验值，即**约 15,000 个输入 token**。按每个中文字符 0.6 token 的经验值，再加上 JSON 和 ID 开销，中文 JSON 输出估计为 3,500 到 4,500 token。
 
-如果所有输入都按缓存未命中计费，输入约 $0.0046，输出约 $0.0010 到 $0.0013，总计约 $0.0056 到 $0.0059。当大量重复的 system prompt 命中 DeepSeek 自动尽力而为缓存时，更现实的低值约为 $0.002 到 $0.003。完整翻译这段演讲的实用估算是 **$0.002 到 $0.006 USD，约 ¥0.02 到 ¥0.04**。
+如果所有输入都按缓存未命中计费，输入约 $0.0021，输出约 $0.0010 到 $0.0013，总计约 $0.0031 到 $0.0034。当大量重复的 system prompt 命中 DeepSeek 自动尽力而为缓存时，更现实的低值约为 $0.001 到 $0.002。完整翻译这段演讲的实用估算是 **$0.001 到 $0.004 USD，约 ¥0.01 到 ¥0.03**。
 
 翻译是延迟按需和渐进式的。已缓存的分段会复用，只有滚动到并请求的字幕行才会发起调用。重试、服务商行为和价格变化都可能增加最终成本。
 

background.js

@@ -1448,8 +1448,8 @@ async function getTranslationBaseRules(targetLanguage) {
 
 function validateTranscriptBatchRequest(content) {
   const segments = content?.segments;
-  if (!Array.isArray(segments) || segments.length < 1 || segments.length > 4) {
-    throw new Error("Transcript translation requires 1 to 4 segments");
+  if (!Array.isArray(segments) || segments.length < 1 || segments.length > 8) {
+    throw new Error("Transcript translation requires 1 to 8 segments");
   }
 
   const seenIds = new Set();
@@ -1563,7 +1563,7 @@ async function handleTranslateContent(
     const userContent = JSON.stringify({ segments: sourceSegments });
     const translationOptions = {
       temperature: 0.2,
-      maxTokens: 1536,
+      maxTokens: 4096,
       responseFormat: { type: "json_object" },
     };
     let result = await callAiTranslation(

prompts/translation.md

@@ -27,7 +27,7 @@ TRANSLATION RULES (follow strictly):
 
 ## Transcript batch translation
 
-Input is a JSON object with 1 to 4 complete semantic transcript segments. Each
+Input is a JSON object with 1 to 8 complete semantic transcript segments. Each
 segment has a stable `id` and source-language `text`.
 
 ```

sidepanel.js

@@ -39,6 +39,7 @@ let transcriptScrollObserver = null;
 // Stable keys include the video, source mode, language, and semantic segment ID.
 let transcriptParagraphCache = new Map();
 const TRANSLATION_MESSAGE_TIMEOUT_MS = 130_000;
+const TRANSLATION_BATCH_SIZE = 8;
 
 /**
  * Prevent a stopped service worker or dead message channel from leaving the
@@ -1989,7 +1990,7 @@ function retryTranslationSegment(index, generation) {
 }
 
 /**
- * Renders immediately, translates the first small batch, then observes the
+ * Renders immediately, translates the first visible batch, then observes the
  * remaining rows. Batches are sequential so the provider is never flooded.
  */
 async function translateTranscript() {
@@ -2011,7 +2012,7 @@ async function translateTranscript() {
     if (processing || queue.length === 0 || generation !== translationGeneration)
       return;
     processing = true;
-    const indices = queue.splice(0, 3);
+    const indices = queue.splice(0, TRANSLATION_BATCH_SIZE);
     indices.forEach((index) => queued.delete(index));
     try {
       await requestTranscriptTranslationBatch(
@@ -2061,7 +2062,7 @@ async function translateTranscript() {
 
   rows.forEach((row, index) => {
     if (!row.classList.contains("translated")) transcriptScrollObserver.observe(row);
-    if (index < 3) enqueue(index);
+    if (index < TRANSLATION_BATCH_SIZE) enqueue(index);
   });
 }
 

tests/release.test.js

@@ -101,15 +101,15 @@ test("release copy documents current scope without em dashes", () => {
   assert.match(readme, /api-docs\.deepseek\.com\/guides\/kv_cache/i);
   assert.match(readme, /\$0\.0028[\s\S]*\$0\.14[\s\S]*\$0\.28/);
   assert.match(readme, /2,935 spoken English words/i);
-  assert.match(readme, /about 32,600 input tokens/i);
-  assert.match(readme, /\$0\.002[^\n]*\$0\.006 USD/i);
+  assert.match(readme, /about 15,000 input tokens/i);
+  assert.match(readme, /\$0\.001[^\n]*\$0\.004 USD/i);
   assert.match(chineseReadme, /api-docs\.deepseek\.com\/quick_start\/pricing/i);
   assert.match(chineseReadme, /api-docs\.deepseek\.com\/quick_start\/token_usage/i);
   assert.match(chineseReadme, /api-docs\.deepseek\.com\/guides\/kv_cache/i);
   assert.match(chineseReadme, /\u00a50\.02[\s\S]*\u00a51[\s\S]*\u00a52/);
   assert.match(chineseReadme, /2,935 \u4e2a\u82f1\u6587\u53e3\u8bed\u8bcd/);
-  assert.match(chineseReadme, /\u7ea6 32,600 \u4e2a\u8f93\u5165 token/);
-  assert.match(chineseReadme, /\$0\.002[^\n]*\$0\.006 USD/);
+  assert.match(chineseReadme, /\u7ea6 15,000 \u4e2a\u8f93\u5165 token/);
+  assert.match(chineseReadme, /\$0\.001[^\n]*\$0\.004 USD/);
   assert.match(chineseReadme, /dash\.supadata\.ai\/auth\/sign-up/i);
   assert.match(chineseReadme, /platform\.deepseek\.com\/api_keys/i);
   assert.match(readme, /^### The Digest button is missing on a YouTube video$/m);

tests/translation.test.js

@@ -173,6 +173,7 @@ test("Transcript header exposes and wires Original, Chinese, and bilingual modes
   assert.match(html, /data-transcript-mode="bilingual"[\s\S]*?>\u53cc\u8bed</);
   assert.match(js, /handleTranscriptModeChange\(button\.dataset\.transcriptMode\)/);
   assert.match(js, /contentType: "transcriptBatch"/);
+  assert.match(js, /TRANSLATION_BATCH_SIZE = 8/);
   assert.doesNotMatch(js, /English \+ Chinese/);
   assert.match(js, /Original \(\$\{language\}\)/);
 });
@@ -315,7 +316,19 @@ test("background rejects unsupported language fallthrough and malformed batches"
   assert.match(source, /targetLanguage !== "zh"/);
   assert.throws(
     () => validateTranscriptBatchRequest({ segments: [] }),
-    /1 to 4 segments/,
+    /1 to 8 segments/,
+  );
+  const eight = Array.from({ length: 8 }, (_, index) => ({
+    id: `segment-${index}-0`,
+    text: "Hello.",
+  }));
+  assert.equal(validateTranscriptBatchRequest({ segments: eight }).length, 8);
+  assert.throws(
+    () =>
+      validateTranscriptBatchRequest({
+        segments: [...eight, { id: "segment-8-0", text: "Too many." }],
+      }),
+    /1 to 8 segments/,
   );
   assert.throws(
     () =>
@@ -544,7 +557,7 @@ test("DeepSeek retries one empty transcript JSON response without response_forma
   assert.equal(requests.length, 2);
   assert.deepEqual(requests[0].response_format, { type: "json_object" });
   assert.equal(Object.hasOwn(requests[1], "response_format"), false);
-  assert.equal(requests[0].max_tokens, 1536);
+  assert.equal(requests[0].max_tokens, 4096);
 });
 
 test("translation message watchdog rejects, clears its timer, and ignores late replies", async () => {