@aj-archipelago/cortex 1.1.23 → 1.1.24

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aj-archipelago/cortex",
-  "version": "1.1.23",
+  "version": "1.1.24",
   "description": "Cortex is a GraphQL API for AI. It provides a simple, extensible interface for using AI services from OpenAI, Azure and others.",
   "private": false,
   "repository": {

package/pathways/translate_subtitle.js

@@ -1,6 +1,7 @@
 import subsrt from "subsrt";
 import logger from "../lib/logger.js";
 import { callPathway } from "../lib/pathwayTools.js";
+import { publishRequestProgress } from "../lib/redisSubscription.js";
 
 function preprocessStr(str) {
   try {
@@ -18,64 +19,25 @@ function preprocessStr(str) {
   }
 }
 
-function getContextLines(captions, startIndex, direction, wordLimit = 100) {
-  let context = "";
-  let wordCount = 0;
-  let i = startIndex;
-
-  while (i >= 0 && i < captions.length && wordCount < wordLimit) {
-    const words = captions[i].content.split(/\s+/);
-    if (wordCount + words.length <= wordLimit) {
-      context =
-        direction === "prev"
-          ? captions[i].content + " " + context
-          : context + " " + captions[i].content;
-      wordCount += words.length;
-    } else {
-      const remainingWords = wordLimit - wordCount;
-      const partialContent =
-        direction === "prev"
-          ? words.slice(-remainingWords).join(" ")
-          : words.slice(0, remainingWords).join(" ");
-      context =
-        direction === "prev"
-          ? partialContent + " " + context
-          : context + " " + partialContent;
-      break;
-    }
-    i += direction === "prev" ? -1 : 1;
-  }
-
-  return context.trim();
-}
-
-async function processBatch(batch, args, captions, batchStartIndex) {
+async function processBatch(batch, args) {
   const batchText = batch
     .map((caption, index) => `LINE#${index + 1}: ${caption.content}`)
     .join("\n");
-  const prevLines = getContextLines(captions, batchStartIndex - 1, "prev");
-  const nextLines = getContextLines(
-    captions,
-    batchStartIndex + batch.length,
-    "next"
-  );
 
   const translatedText = await callPathway("translate_subtitle_helper", {
     ...args,
     text: batchText,
-    prevLines,
-    nextLines,
     async: false,
   });
 
   // Remove LINE# and LINE() labels
-    const translatedLines = translatedText.split("\n");
+  const translatedLines = translatedText.split("\n");
     translatedLines.forEach((line, i) => {
     translatedLines[i] = line.replace(/^LINE#\d+:\s*/, "").trim();
-    });
+  });
   //make sure translatedLines.length===batch.length
   if (translatedLines.length < batch.length) {
-    const emptyLines = Array(batch.length - translatedLines.length).fill("");
+    const emptyLines = Array(batch.length - translatedLines.length).fill("-");
     translatedLines.push(...emptyLines);
   } else if (translatedLines.length > batch.length) {
     //first remove the empty lines
@@ -88,7 +50,7 @@ async function processBatch(batch, args, captions, batchStartIndex) {
         mergedLines.unshift(lastLine);
         translatedLines.splice(batch.length - 1, translatedLines.length - batch.length + 1, mergedLines.join(" "));
     }else {
-        const emptyLines = Array(batch.length - translatedLines.length).fill("");
+        const emptyLines = Array(batch.length - translatedLines.length).fill("-");
         translatedLines.push(...emptyLines);
     }
   }
@@ -114,7 +76,7 @@ async function processBatch(batch, args, captions, batchStartIndex) {
   }));
 }
 
-async function myResolver(args) {
+async function myResolver(args, requestId) {
   try {
     const { text, format } = args;
     const captions = subsrt.parse(preprocessStr(text), {
@@ -123,11 +85,30 @@ async function myResolver(args) {
       eol: "\n",
     });
     const maxLineCount = 100;
-    const maxWordCount = 1000;
+    const maxWordCount = 300;
     let translatedCaptions = [];
     let currentBatch = [];
     let currentWordCount = 0;
-    let batchStartIndex = 0;
+
+    const totalCount = captions.length;
+    let completedCount = 0;
+
+    const sendProgress = () => {
+      if (completedCount >= totalCount) return;
+      if(!requestId) {
+        logger.warn(`No requestId found for progress update`);
+        return;
+      }
+
+      const progress = completedCount / totalCount;
+      logger.info(`Progress for ${requestId}: ${progress}`);
+
+      publishRequestProgress({
+        requestId,
+        progress,
+        data: null,
+      });
+    };
 
     for (let i = 0; i < captions.length; i++) {
       const caption = captions[i];
@@ -137,16 +118,15 @@ async function myResolver(args) {
           currentBatch.length >= maxLineCount) &&
         currentBatch.length > 0
       ) {
+        completedCount=i;
+        sendProgress();
         const translatedBatch = await processBatch(
           currentBatch,
           args,
-          captions,
-          batchStartIndex
         );
         translatedCaptions = translatedCaptions.concat(translatedBatch);
         currentBatch = [];
         currentWordCount = 0;
-        batchStartIndex = i;
       }
       currentBatch.push(caption);
       currentWordCount += captionWordCount;
@@ -156,8 +136,6 @@ async function myResolver(args) {
       const translatedBatch = await processBatch(
         currentBatch,
         args,
-        captions,
-        batchStartIndex
       );
       translatedCaptions = translatedCaptions.concat(translatedBatch);
     }
@@ -171,11 +149,11 @@ async function myResolver(args) {
         .trim() + "\n"
     );
   } catch (e) {
-    logger.error(
-      `An error occurred in subtitle translation, trying direct translation next: ${e}`
+    logger.warn(
+      `${e} - could be that there are no subtitles, so attempting block translation.`
     );
     try {
-      return await callPathway("translate_gpt4", {...args, async: false});
+      return await callPathway("translate_gpt4_omni", {...args, async: false});
     } catch (e) {
       logger.error(`An error occurred in subtitle translation: ${e}`);
       return "";
@@ -191,11 +169,13 @@ export default {
     prevLines: ``,
     nextLines: ``,
   },
-  inputChunkSize: 500,
+  useInputChunking: false,
   model: "oai-gpt4o",
   enableDuplicateRequests: false,
   timeout: 3600,
-  executePathway: async ({ args }) => {
-    return await myResolver(args);
+  executePathway: async (executePathwayArgs) => {
+    const { args } = executePathwayArgs;
+    const requestId = executePathwayArgs?.resolver?.requestId;
+    return await myResolver(args, requestId);
   },
 };

package/pathways/translate_subtitle_helper.js

@@ -1,6 +1,5 @@
 import { Prompt } from '../server/prompt.js';
 
-
 export default {
     prompt: [
         new Prompt({
@@ -8,11 +7,25 @@ export default {
             {
                 role: "system",
                 content: 
-                `Assistant is a highly skilled multilingual translator for a prestigious news agency. When the user posts any text in any language, assistant will create a translation of that text in {{to}}. User will most probably produce previous and next lines for context with "PreviousLines" and "NextLines" labels, and you are asked to translate current lines one by one in given sequence with "CurrentLines" label. CurrentLines might have numbered labels as LINE#{lineNo} e.g. LINE#1, LINE#2. If currentline is a word only translate that word. You must keep input and output number of lines same, so do not merge translation of lines, single line must always map to single line. Assistant's output translated number of lines must always be equal to the input number of currentlines. For output, Assistant will produce only the translated text, ignore all LINE#{lineNo} and "CurrentLines" labels, and give no additional notes or commentary.`,
+`Expert translator: Convert ALL text to {{to}}. Unbreakable rules:
+
+1. Translate EVERY SINGLE LINE. Zero exceptions.
+2. Output MUST have EXACTLY the same line count as input.
+3. One input line = One output line. Always.
+4. Only translations. Nothing extra.
+5. Non-translatable stays unchanged.
+6. Keep all formatting and characters.
+7. Prefix: "LINE#lineNumber:".
+8. Untranslatable: Copy as-is with prefix.
+9. Internal checks: Verify line count and content after each line.
+10. Final verification: Recount, check numbering, confirm content, cross-check with input.
+
+Translate ALL lines. Constant vigilance. Exhaustive final cross-check.`
             },
             {
                 role: "user",
-                content: `"PreviousLines":\n{{{prevLine}}}\n\n"CurrentLines":\n{{{text}}}\n"NextLines":\n{{{nextLine}}}\n\n`,
+                // content: `"PreviousLines":\n{{{prevLine}}}\n\n"CurrentLines":\n{{{text}}}\n"NextLines":\n{{{nextLine}}}\n\n`,
+                content: `{{{text}}}`,
             },
             ],
         }),
@@ -24,8 +37,8 @@ export default {
         prevLine: ``,
         nextLine: ``,
     },
-    inputChunkSize: 500,
+    useInputChunking: false,
     model: 'oai-gpt4o',
     enableDuplicateRequests: false,
-
+    timeout: 3600,
 }

package/server/pathwayResolver.js

@@ -206,7 +206,7 @@ class PathwayResolver {
 
     async executePathway(args) {
         if (this.pathway.executePathway && typeof this.pathway.executePathway === 'function') {
-            return await this.pathway.executePathway({ args, runAllPrompts: this.promptAndParse.bind(this) });
+            return await this.pathway.executePathway({ args, runAllPrompts: this.promptAndParse.bind(this), resolver: this });
         }
         else {
             return await this.promptAndParse(args);

package/server/plugins/openAiWhisperPlugin.js

@@ -37,14 +37,13 @@ class OpenAIWhisperPlugin extends ModelPlugin {
                 chunks.push(chunk);
 
                 const { language, responseFormat } = parameters;
-                cortexRequest.url = this.requestUrl(text);
                 const params = {};
                 const { modelPromptText } = this.getCompiledPrompt(text, parameters, prompt);
                 const response_format = responseFormat || 'text';
 
                 const formData = new FormData();
                 formData.append('file', fs.createReadStream(chunk));
-                formData.append('model', this.model.params.model);
+                formData.append('model', cortexRequest.params.model);
                 formData.append('response_format', response_format);
                 language && formData.append('language', language);
                 modelPromptText && formData.append('prompt', modelPromptText);

package/tests/main.test.js

@@ -405,6 +405,13 @@ Aseel is mommy
 });
 
 test('test translate_srt endpoint with long srt file', async t => {
+    t.timeout(400000);
     const text = fs.readFileSync(path.join(__dirname, 'sublong.srt'), 'utf8');
     await testTranslateSrt(t, text, 'English');
+});
+
+test('test translate_srt endpoint with horizontal srt file', async t => {
+    t.timeout(400000);
+    const text = fs.readFileSync(path.join(__dirname, 'subhorizontal.srt'), 'utf8');
+    await testTranslateSrt(t, text, 'Turkish');
 });