@aj-archipelago/cortex 1.3.22 → 1.3.24

package/README.md

@@ -561,6 +561,70 @@ Each model configuration can include:
 }
 ```
 
+### API Compatibility
+
+Cortex provides OpenAI-compatible REST endpoints that allow you to use various models through a standardized interface. When `enableRestEndpoints` is set to `true`, Cortex exposes the following endpoints:
+
+- `/v1/models`: List available models
+- `/v1/chat/completions`: Chat completion endpoint
+- `/v1/completions`: Text completion endpoint
+
+This means you can use Cortex with any client library or tool that supports the OpenAI API format. For example:
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    base_url="http://localhost:4000/v1",  # Point to your Cortex server
+    api_key="your-key"  # If you have configured cortexApiKeys
+)
+
+response = client.chat.completions.create(
+    model="gpt-4",  # Or any model configured in Cortex
+    messages=[{"role": "user", "content": "Hello!"}]
+)
+```
+
+#### Ollama Integration
+
+Cortex includes built-in support for Ollama models through its OpenAI-compatible REST interface. When `ollamaUrl` is configured in your settings, Cortex will:
+1. Automatically discover and expose all available Ollama models through the `/v1/models` endpoint with an "ollama-" prefix
+2. Route any requests using an "ollama-" prefixed model to the appropriate Ollama endpoint
+
+To enable Ollama support, add the following to your configuration:
+
+```json
+{
+    "enableRestEndpoints": true,
+    "ollamaUrl": "http://localhost:11434"  // or your Ollama server URL
+}
+```
+
+You can then use any Ollama model through the standard OpenAI-compatible endpoints:
+
+```bash
+# List available models (will include Ollama models with "ollama-" prefix)
+curl http://localhost:4000/v1/models
+
+# Use an Ollama model for chat
+curl http://localhost:4000/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "ollama-llama2",
+    "messages": [{"role": "user", "content": "Hello!"}]
+  }'
+
+# Use an Ollama model for completions
+curl http://localhost:4000/v1/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "ollama-codellama",
+    "prompt": "Write a function that"
+  }'
+```
+
+This integration allows you to seamlessly use local Ollama models alongside cloud-based models through a single, consistent interface.
+
 ### Other Configuration Properties
 
 The following properties can be configured through environment variables or the configuration file:

package/config.js

@@ -85,6 +85,11 @@ var config = convict({
         default: false,
         env: 'CORTEX_ENABLE_REST'
     },
+    ollamaUrl: {
+        format: String,
+        default: 'http://127.0.0.1:11434',
+        env: 'OLLAMA_URL'
+    },
     entityConstants: {
         format: Object,
         default: {
@@ -281,7 +286,27 @@ var config = convict({
                 "headers": {
                     "Content-Type": "application/json"
                 },
-            }
+            },
+            "ollama-chat": {
+                "type": "OLLAMA-CHAT",
+                "url": "{{ollamaUrl}}/api/chat",
+                "headers": {
+                  "Content-Type": "application/json"
+                },
+                "requestsPerSecond": 10,
+                "maxTokenLength": 131072,
+                "supportsStreaming": true
+            },
+            "ollama-completion": {
+                "type": "OLLAMA-COMPLETION",
+                "url": "{{ollamaUrl}}/api/generate",
+                "headers": {
+                  "Content-Type": "application/json"
+                },
+                "requestsPerSecond": 10,
+                "maxTokenLength": 131072,
+                "supportsStreaming": true
+            },
         },
         env: 'CORTEX_MODELS'
     },

package/helper-apps/cortex-realtime-voice-server/src/realtime/client.ts

@@ -355,14 +355,19 @@ export class RealtimeVoiceClient extends EventEmitter implements TypedEmitter {
     if (!this.isConnected) {
       throw new Error('Not connected');
     }
+    
+    // Create a new config object without custom_voice_id
+    const { custom_voice_id, ...filteredConfig } = {
+      ...this.sessionConfig,
+      ...sessionConfig
+    };
+
     const message = JSON.stringify({
       event_id: createId(),
       type: 'session.update',
-      session: {
-        ...this.sessionConfig,
-        ...sessionConfig,
-      },
+      session: filteredConfig,
     });
+    
     // No need to log session update messages as they can be noisy
     logger.log('Sending session update message:', message);
     this.ws?.send(message);

package/helper-apps/cortex-realtime-voice-server/src/realtime/realtimeTypes.ts

@@ -46,6 +46,7 @@ export type RealtimeSessionConfig = {
   modalities: Array<Modality>,
   instructions: string,
   voice: Voice,
+  custom_voice_id?: string | null,
   input_audio_format: AudioFormat,
   output_audio_format: AudioFormat,
   input_audio_transcription: null | { model: 'whisper-1' | (string & {}) },

package/lib/util.js

@@ -1,6 +1,6 @@
 import logger from "./logger.js";
 import stream from 'stream';
-import subsrt from 'subsrt';
+import subvibe from '@aj-archipelago/subvibe';
 import os from 'os';
 import http from 'http';
 import https from 'https';
@@ -126,19 +126,9 @@ function convertSrtToText(str) {
 function alignSubtitles(subtitles, format, offsets) {
     const result = [];
 
-    function preprocessStr(str) {
-        try{
-            if(!str) return '';
-            return str.trim().replace(/(\n\n)(?!\n)/g, '\n\n\n');
-        }catch(e){
-            logger.error(`An error occurred in content text preprocessing: ${e}`);
-            return '';
-        }
-    }
-
     function shiftSubtitles(subtitle, shiftOffset) {
-        const captions = subsrt.parse(preprocessStr(subtitle));
-        const resynced = subsrt.resync(captions, { offset: shiftOffset });
+        const captions = subvibe.parse(subtitle);
+        const resynced = subvibe.resync(captions.cues, { offset: shiftOffset });
         return resynced;
     }
 
@@ -146,18 +136,8 @@ function alignSubtitles(subtitles, format, offsets) {
         result.push(...shiftSubtitles(subtitles[i], offsets[i]*1000)); // convert to milliseconds
     }
     
-    try {
-        //if content has needed html style tags, keep them
-        for(const obj of result) {
-            if(obj && obj.content){ 
-                obj.text = obj.content;
-            }
-        }
-    } catch (error) {
-        logger.error(`An error occurred in content text parsing: ${error}`);
-    }
     
-    return subsrt.build(result, { format: format === 'vtt' ? 'vtt' : 'srt' });
+    return subvibe.build(result, format || 'srt');
 }
 
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aj-archipelago/cortex",
-  "version": "1.3.22",
+  "version": "1.3.24",
   "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": {
@@ -33,6 +33,7 @@
   "type": "module",
   "homepage": "https://github.com/aj-archipelago/cortex#readme",
   "dependencies": {
+    "@aj-archipelago/subvibe": "^1.0.5",
     "@apollo/server": "^4.7.3",
     "@apollo/server-plugin-response-cache": "^4.1.2",
     "@apollo/utils.keyvadapter": "^3.0.0",
@@ -63,7 +64,6 @@
     "ioredis": "^5.3.1",
     "keyv": "^4.5.2",
     "mime-types": "^2.1.35",
-    "subsrt": "^1.1.1",
     "uuid": "^9.0.0",
     "winston": "^3.11.0",
     "ws": "^8.12.0"
@@ -86,5 +86,8 @@
       "dotenv/config"
     ],
     "concurrency": 1
+  },
+  "overrides": {
+    "whatwg-url": "^12.0.0"
   }
 }

package/pathways/system/entity/sys_generator_memory.js

@@ -18,7 +18,7 @@ export default {
         const { aiStyle, AI_STYLE_ANTHROPIC, AI_STYLE_OPENAI } = args;
         const styleModel = aiStyle === "Anthropic" ? AI_STYLE_ANTHROPIC : AI_STYLE_OPENAI;
 
-        const memoryContext = await callPathway('sys_search_memory', { ...args, section: 'memoryAll', updateContext: true });
+        const memoryContext = await callPathway('sys_search_memory', { ...args, stream: false, section: 'memoryAll', updateContext: true });
         if (memoryContext) {
             const {toolCallId} = addToolCalls(args.chatHistory, "search memory for relevant information", "memory_lookup");
             addToolResults(args.chatHistory, memoryContext, toolCallId);
@@ -26,9 +26,9 @@ export default {
 
         let result;
         if (args.voiceResponse) {
-            result = await callPathway('sys_generator_quick', { ...args, model: styleModel, stream: false });
+            result = await callPathway('sys_generator_quick', { ...args, model: styleModel, stream: false }, resolver);
         } else {
-            result = await callPathway('sys_generator_quick', { ...args, model: styleModel });
+            result = await callPathway('sys_generator_quick', { ...args, model: styleModel }, resolver);
         }
 
         resolver.tool = JSON.stringify({ toolUsed: "memory" });

package/pathways/system/rest_streaming/sys_ollama_chat.js

@@ -0,0 +1,21 @@
+// sys_ollama_chat.js
+// override handler for ollama chat model
+
+import { Prompt } from '../../../server/prompt.js';
+
+export default {
+    prompt:
+    [
+        new Prompt({ messages: [
+            "{{messages}}",
+        ]}),
+    ],
+    inputParameters: {
+        messages: [{ role: '', content: '' }],
+        ollamaModel: '',
+    },
+    model: 'ollama-chat',
+    useInputChunking: false,
+    emulateOpenAIChatModel: 'ollama-chat',
+    timeout: 300,
+}

package/pathways/system/rest_streaming/sys_ollama_completion.js

@@ -0,0 +1,14 @@
+// sys_ollama_completion.js
+// default handler for ollama completion endpoints when REST endpoints are enabled
+
+export default {
+    prompt: `{{text}}`,
+    inputParameters: {
+        text: '',
+        ollamaModel: '',
+    },
+    model: 'ollama-completion',
+    useInputChunking: false,
+    emulateOpenAICompletionModel: 'ollama-completion',
+    timeout: 300,
+}

package/pathways/system/rest_streaming/sys_openai_chat.js

@@ -13,7 +13,7 @@ export default {
     inputParameters: {
         messages: [],
     },
-    model: 'oai-gpturbo',
+    model: 'oai-gpt4o',
     useInputChunking: false,
-    emulateOpenAIChatModel: 'gpt-3.5-turbo',
+    emulateOpenAIChatModel: 'gpt-4o',
 }

package/pathways/transcribe_gemini.js

@@ -5,36 +5,124 @@ import { Prompt } from "../server/prompt.js";
 
 const OFFSET_CHUNK = 500; //seconds of each chunk offset, only used if helper does not provide
 
-
-
 export function convertSrtToVtt(data) {
     if (!data || !data.trim()) {
         return "WEBVTT\n\n";
     }
-    // remove dos newlines
+
+    // If it's already VTT format and has header
+    if (data.trim().startsWith("WEBVTT")) {
+        const lines = data.split("\n");
+        const result = ["WEBVTT", ""]; // Start with header and blank line
+        let currentCue = [];
+
+        for (let i = 0; i < lines.length; i++) {
+            const line = lines[i].trim();
+
+            // Skip empty lines and the WEBVTT header
+            if (!line || line === "WEBVTT") {
+                continue;
+            }
+
+            // If it's a number by itself, it's a cue identifier
+            if (/^\d+$/.test(line)) {
+                // If we have a previous cue, add it with proper spacing
+                if (currentCue.length > 0) {
+                    result.push(currentCue.join("\n"));
+                    result.push(""); // Add blank line between cues
+                    currentCue = [];
+                }
+                currentCue.push(line);
+                continue;
+            }
+
+            // Check for and convert timestamps
+            const fullTimeRegex = /^(\d{2}):(\d{2}):(\d{2})[,.](\d{3})\s*-->\s*(\d{2}):(\d{2}):(\d{2})[,.](\d{3})$/;
+            const shortTimeRegex = /^(\d{2}):(\d{2})[,.](\d{3})\s*-->\s*(\d{2}):(\d{2})[,.](\d{3})$/;
+            const ultraShortTimeRegex = /^(\d{1,2})[.](\d{3})\s*-->\s*(\d{1,2})[.](\d{3})$/;
+
+            const fullMatch = line.match(fullTimeRegex);
+            const shortMatch = line.match(shortTimeRegex);
+            const ultraShortMatch = line.match(ultraShortTimeRegex);
+
+            if (fullMatch) {
+                // Already in correct format, just convert comma to dot
+                const convertedTime = line.replace(/,/g, '.');
+                currentCue.push(convertedTime);
+            } else if (shortMatch) {
+                // Convert MM:SS to HH:MM:SS
+                const convertedTime = `00:${shortMatch[1]}:${shortMatch[2]}.${shortMatch[3]} --> 00:${shortMatch[4]}:${shortMatch[5]}.${shortMatch[6]}`;
+                currentCue.push(convertedTime);
+            } else if (ultraShortMatch) {
+                // Convert SS to HH:MM:SS
+                const convertedTime = `00:00:${ultraShortMatch[1].padStart(2, '0')}.${ultraShortMatch[2]} --> 00:00:${ultraShortMatch[3].padStart(2, '0')}.${ultraShortMatch[4]}`;
+                currentCue.push(convertedTime);
+            } else if (!line.includes('-->')) {
+                // Must be subtitle text
+                currentCue.push(line);
+            }
+        }
+
+        // Add the last cue if there is one
+        if (currentCue.length > 0) {
+            result.push(currentCue.join("\n"));
+            result.push(""); // Add final blank line
+        }
+
+        // Join with newlines and ensure proper ending
+        return result.join("\n") + "\n";
+    }
+
+    // remove dos newlines and trim
     var srt = data.replace(/\r+/g, "");
-    // trim white space start and end
     srt = srt.replace(/^\s+|\s+$/g, "");
 
-    // Convert all timestamps from comma to dot format
-    srt = srt.replace(/(\d{2}:\d{2}:\d{2}),(\d{3})/g, "$1.$2");
-
-    // Add blank lines before sequence numbers that are followed by timecodes
-    srt = srt.replace(/(\n)(\d+)\n(\d{2}:\d{2}:\d{2}[,.])/g, "$1\n$2\n$3");
-
-    // get cues
-    var cuelist = srt.split("\n\n");
-    var result = "";
-    if (cuelist.length > 0) {
-        result += "WEBVTT\n\n";
-        for (var i = 0; i < cuelist.length; i = i + 1) {
-            const cue = convertSrtCue(cuelist[i]);
-            // Only add non-empty cues
-            if (cue) {
-                result += cue;
-            }
+    // Split into cues and filter out empty ones
+    var cuelist = srt.split("\n\n").filter(cue => cue.trim());
+
+    // Always add WEBVTT header
+    var result = "WEBVTT\n\n";
+
+    // Convert each cue to VTT format
+    for (const cue of cuelist) {
+        const lines = cue.split("\n").map(line => line.trim()).filter(line => line);
+        if (lines.length < 2) continue;
+
+        let output = [];
+        
+        // Handle cue identifier
+        if (/^\d+$/.test(lines[0])) {
+            output.push(lines[0]);
+            lines.shift();
+        }
+
+        // Handle timestamp line
+        const timeLine = lines[0];
+        const fullTimeRegex = /^(\d{2}):(\d{2}):(\d{2})[,.](\d{3})\s*-->\s*(\d{2}):(\d{2}):(\d{2})[,.](\d{3})$/;
+        const shortTimeRegex = /^(\d{2}):(\d{2})[,.](\d{3})\s*-->\s*(\d{2}):(\d{2})[,.](\d{3})$/;
+        const ultraShortTimeRegex = /^(\d{1,2})[.](\d{3})\s*-->\s*(\d{1,2})[.](\d{3})$/;
+
+        const fullMatch = timeLine.match(fullTimeRegex);
+        const shortMatch = timeLine.match(shortTimeRegex);
+        const ultraShortMatch = timeLine.match(ultraShortTimeRegex);
+
+        if (fullMatch) {
+            output.push(timeLine.replace(/,/g, '.'));
+        } else if (shortMatch) {
+            output.push(`00:${shortMatch[1]}:${shortMatch[2]}.${shortMatch[3]} --> 00:${shortMatch[4]}:${shortMatch[5]}.${shortMatch[6]}`);
+        } else if (ultraShortMatch) {
+            output.push(`00:00:${ultraShortMatch[1].padStart(2, '0')}.${ultraShortMatch[2]} --> 00:00:${ultraShortMatch[3].padStart(2, '0')}.${ultraShortMatch[4]}`);
+        } else {
+            continue; // Invalid timestamp format
         }
+
+        // Add remaining lines as subtitle text
+        output.push(...lines.slice(1));
+        
+        // Add the cue to result
+        result += output.join("\n") + "\n\n";
     }
+
     return result;
 }
 
@@ -42,18 +130,20 @@ function convertSrtCue(caption) {
     if (!caption || !caption.trim()) {
         return "";
     }
-    // remove all html tags for security reasons
-    //srt = srt.replace(/<[a-zA-Z\/][^>]*>/g, '');
+
     var cue = "";
     var s = caption.split(/\n/);
-    // concatenate muilt-line string separated in array into one
+
+    // concatenate multi-line string separated in array into one
     while (s.length > 3) {
         for (var i = 3; i < s.length; i++) {
             s[2] += "\n" + s[i];
         }
         s.splice(3, s.length - 3);
     }
+
     var line = 0;
+
     // detect identifier
     if (
         s[0] &&
@@ -67,10 +157,11 @@ function convertSrtCue(caption) {
             line += 1;
         }
     }
+
     // get time strings
     if (s[line] && s[line].match(/\d+:\d+:\d+/)) {
         // convert time string
-        var m = s[1].match(
+        var m = s[line].match(
             /(\d{2}):(\d{2}):(\d{2})[,.](\d{3})\s*--?>\s*(\d{2}):(\d{2}):(\d{2})[,.](\d{3})/,
         );
         if (m) {
@@ -93,17 +184,43 @@ function convertSrtCue(caption) {
                 "\n";
             line += 1;
         } else {
-            // Unrecognized timestring
-            return "";
+            // Try alternate timestamp format
+            m = s[line].match(
+                /(\d{2}):(\d{2})\.(\d{3})\s*--?>\s*(\d{2}):(\d{2})\.(\d{3})/,
+            );
+            if (m) {
+                // Convert to full timestamp format
+                cue +=
+                    "00:" +
+                    m[1] +
+                    ":" +
+                    m[2] +
+                    "." +
+                    m[3] +
+                    " --> " +
+                    "00:" +
+                    m[4] +
+                    ":" +
+                    m[5] +
+                    "." +
+                    m[6] +
+                    "\n";
+                line += 1;
+            } else {
+                // Unrecognized timestring
+                return "";
+            }
         }
     } else {
         // file format error or comment lines
         return "";
     }
+
     // get cue text
     if (s[line]) {
         cue += s[line] + "\n\n";
     }
+
     return cue;
 }
 
@@ -112,48 +229,59 @@ export function detectSubtitleFormat(text) {
     const cleanText = text.replace(/\r+/g, "").trim();
     const lines = cleanText.split("\n");
 
-    // Check if it's VTT format
+    // Check if it's VTT format - be more lenient with the header
     if (lines[0]?.trim() === "WEBVTT") {
         return "vtt";
     }
 
-    // Check if it's SRT format
-    // SRT files have a specific pattern:
-    // 1. Numeric index
-    // 2. Timestamp in format: 00:00:00,000 --> 00:00:00,000
-    // 3. Subtitle text
-    // 4. Blank line
-    const timeRegex =
-        /(\d{2}:\d{2}:\d{2})[,.](\d{3})\s*-->\s*(\d{2}:\d{2}:\d{2})[,.](\d{3})/;
+    // Define regex patterns for timestamp formats
+    const srtTimeRegex =
+        /(\d{2}:\d{2}:\d{2})[,.]\d{3}\s*-->\s*(\d{2}:\d{2}:\d{2})[,.]\d{3}/;
+    const vttTimeRegex =
+        /(?:\d{2}:)?(\d{1,2})[.]\d{3}\s*-->\s*(?:\d{2}:)?(\d{1,2})[.]\d{3}/;
 
-    let hasValidStructure = false;
-    let index = 1;
+    let hasSrtTimestamps = false;
+    let hasVttTimestamps = false;
+    let hasSequentialNumbers = false;
+    let lastNumber = 0;
 
-    // Check first few entries to confirm SRT structure
+    // Look through first few lines to detect patterns
     for (let i = 0; i < Math.min(lines.length, 12); i++) {
         const line = lines[i]?.trim();
         if (!line) continue;
 
-        // Check if line is a number matching our expected index
-        if (line === index.toString()) {
-            // Look ahead for timestamp
-            const nextLine = lines[i + 1]?.trim();
-            if (nextLine && timeRegex.test(nextLine)) {
-                hasValidStructure = true;
-                index++;
-                i++; // Skip timestamp line since we've verified it
+        // Check for timestamps
+        if (srtTimeRegex.test(line)) {
+            hasSrtTimestamps = true;
+        }
+        if (vttTimeRegex.test(line)) {
+            hasVttTimestamps = true;
+        }
+
+        // Check for sequential numbers
+        const numberMatch = line.match(/^(\d+)$/);
+        if (numberMatch) {
+            const num = parseInt(numberMatch[1]);
+            if (lastNumber === 0 || num === lastNumber + 1) {
+                hasSequentialNumbers = true;
+                lastNumber = num;
             }
         }
     }
 
-    if (hasValidStructure) {
+    // If it has SRT-style timestamps (HH:MM:SS), it's SRT
+    if (hasSrtTimestamps && hasSequentialNumbers) {
         return "srt";
     }
 
+    // If it has VTT-style timestamps (MM:SS) or WEBVTT header, it's VTT
+    if (hasVttTimestamps) {
+        return "vtt";
+    }
+
     return null;
 }
 
-
 export default {
     prompt:
     [
@@ -203,7 +331,6 @@ export default {
             const progress = (completedCount + partialRatio) / totalCount;
             logger.info(`Progress for ${requestId}: ${progress}`);
 
-            console.log(`Progress for ${requestId}: ${progress}`);
             publishRequestProgress({
                 requestId,
                 progress,
@@ -290,7 +417,7 @@ WEBVTT
 Hello World2!
 
 2
-00:05.344 --> 00:00:08.809
+00:00:05.344 --> 00:00:08.809
 Being AI is also great!
 
 - If asked text format, e.g.:
@@ -327,6 +454,7 @@ Even a single newline or space can cause the response to be rejected. You must f
             return messages;
         }
 
+
         const processChunksParallel = async (chunks, args) => {
             try {
                 const chunkPromises = chunks.map(async (chunk, index) => ({
@@ -338,8 +466,6 @@ Even a single newline or space can cause the response to be rejected. You must f
                     })
                 }));
         
-                // const results = await Promise.all(chunkPromises);
-
                 const results = await Promise.all(
                 chunkPromises.map(promise => 
                     promise.then(result => {
@@ -374,6 +500,8 @@ Even a single newline or space can cause the response to be rejected. You must f
         
         if (['srt','vtt'].includes(responseFormat) || wordTimestamped) { // align subtitles for formats
 
+            
+
             // convert as gemini output is unstable
             for(let i = 0; i < result.length; i++) {
                 try{

package/server/modelExecutor.js

@@ -28,6 +28,8 @@ import NeuralSpacePlugin from './plugins/neuralSpacePlugin.js';
 import RunwareAiPlugin from './plugins/runwareAiPlugin.js';
 import ReplicateApiPlugin from './plugins/replicateApiPlugin.js';
 import AzureVideoTranslatePlugin from './plugins/azureVideoTranslatePlugin.js';
+import OllamaChatPlugin from './plugins/ollamaChatPlugin.js';
+import OllamaCompletionPlugin from './plugins/ollamaCompletionPlugin.js';
 
 class ModelExecutor {
     constructor(pathway, model) {
@@ -116,6 +118,12 @@ class ModelExecutor {
             case 'AZURE-VIDEO-TRANSLATE':
                 plugin = new AzureVideoTranslatePlugin(pathway, model);
                 break;
+            case 'OLLAMA-CHAT':
+                plugin = new OllamaChatPlugin(pathway, model);
+                break;
+            case 'OLLAMA-COMPLETION':
+                plugin = new OllamaCompletionPlugin(pathway, model);
+                break;
             default:
                 throw new Error(`Unsupported model type: ${model.type}`);
         }