@aj-archipelago/cortex 1.1.5 → 1.1.7

package/config.js

@@ -118,7 +118,7 @@ var config = convict({
                     "api-key": "{{AZURE_COGNITIVE_API_KEY}}",
                     "Content-Type": "application/json"
                 },
-                "requestsPerSecond": 6
+                "requestsPerSecond": 10
             },
             "oai-embeddings": {
                 "type": "OPENAI-EMBEDDINGS",
@@ -146,6 +146,16 @@ var config = convict({
                 "maxTokenLength": 128000,
                 "supportsStreaming": true
             },
+            "azure-bing": {
+                "type": "AZURE-BING",
+                "url": "https://api.bing.microsoft.com/v7.0/search",
+                "headers": {
+                    "Ocp-Apim-Subscription-Key": "{{AZURE_BING_KEY}}",
+                    "Content-Type": "application/json"
+                },
+                "requestsPerSecond": 10,
+                "maxTokenLength": 200000
+            },
         },
         env: 'CORTEX_MODELS'
     },

package/helper-apps/cortex-whisper-wrapper/app.py

@@ -38,9 +38,14 @@ def transcribe(params):
     if 'word_timestamps' in params: #parse as bool
         word_timestamps = False if params['word_timestamps'] == 'False' else True
 
+    decode_options = {}
+    if 'language' in params:
+        decode_options["language"] = params["language"]
+        print(f"Transcription language set as {decode_options['language']}")
+
     print(f"Transcribing file {fileurl} with word_timestamps={word_timestamps}")
     start_time = time.time()
-    result = model.transcribe(fileurl, word_timestamps=word_timestamps)
+    result = model.transcribe(fileurl, word_timestamps=word_timestamps, **decode_options)
     end_time = time.time()
     execution_time = end_time - start_time
     print("Transcribe execution time:", execution_time, "seconds")

package/lib/cortexRequest.js

@@ -11,6 +11,7 @@ class CortexRequest {
         this._pathwayResolver = pathwayResolver || {};
         this._selectedEndpoint = selectedEndpoint || {};
         this._stream = stream || false;
+        this._method = 'POST';
 
         if (this._pathwayResolver) {
             this._model = this._pathwayResolver.model;
@@ -41,6 +42,15 @@ class CortexRequest {
         this._url = value;
     }
 
+    // method getter and setter
+    get method() {
+        return this._method;
+    }
+
+    set method(value) {
+        this._method = value;
+    }
+
     // data getter and setter
     get data() {
         return this._data;

package/lib/encodeCache.js

@@ -0,0 +1,38 @@
+import { encode as gpt3Encode, decode as gpt3Decode } from 'gpt-3-encoder';
+import { FastLRUCache } from './fastLruCache.js';
+
+class EncodeCache {
+    constructor() {
+        this.encodeCache = new FastLRUCache(1000);
+        this.decodeCache = new FastLRUCache(100); // we don't use decode nearly as much
+    }
+    
+    encode(value) {
+        if (this.encodeCache.get(value) !== -1) {
+            return this.encodeCache.get(value);
+        }
+        const encoded = gpt3Encode(value);
+        this.encodeCache.put(value, encoded);
+        return encoded;
+    }
+
+    decode(value) {
+        if (this.decodeCache.get(value) !== -1) {
+            return this.decodeCache.get(value);
+        }
+        const decoded = gpt3Decode(value);
+        this.decodeCache.put(value, decoded);
+        if (this.encodeCache.get(decoded) === -1) {
+            this.encodeCache.put(decoded, value);
+        }
+        return decoded;
+    }
+}
+
+// Create one instance of the cache
+const cache = new EncodeCache();
+
+// Make sure the instance is bound to the methods, so 
+// references to 'this' are correct
+export const encode = cache.encode.bind(cache);
+export const decode = cache.decode.bind(cache);

package/lib/fastLruCache.js

@@ -0,0 +1,82 @@
+// This class implements a fast O(1) LRU cache using a Map and a doubly linked list.
+
+class Node {
+    constructor(key, value) {
+        this.key = key;
+        this.value = value;
+        this.next = null;
+        this.prev = null;
+    }
+}
+
+class FastLRUCache {
+    constructor(capacity) {
+        this.capacity = capacity;
+        this.cache = new Map();
+        this.head = null;
+        this.tail = null;
+    }
+
+    get(key) {
+        if (!this.cache.has(key)) {
+            return -1;
+        }
+        const node = this.cache.get(key);
+        this.moveToEnd(node);
+        return node.value;
+    }
+
+    put(key, value) {
+        if (this.cache.has(key)) {
+            const node = this.cache.get(key);
+            node.value = value;
+            this.moveToEnd(node);
+        } else {
+            const node = new Node(key, value);
+            if (this.cache.size >= this.capacity) {
+                this.cache.delete(this.head.key);
+                this.shiftHeadToNext();
+            }
+            this.cache.set(key, node);
+            this.addNodeToTail(node);
+        }
+    }
+
+    addNodeToTail(node) {
+        if (!this.tail) {
+            this.head = node;
+            this.tail = node;
+        } else {
+            node.prev = this.tail;
+            this.tail.next = node;
+            this.tail = node;
+        }
+    }
+
+    moveToEnd(node) {
+        if (node === this.tail) {
+            return;
+        }
+        if (node === this.head) {
+            this.shiftHeadToNext();
+        } else {
+            node.prev.next = node.next;
+            node.next.prev = node.prev;
+        }
+        node.prev = this.tail;
+        node.next = null;
+        this.tail.next = node;
+        this.tail = node;
+    }
+
+    shiftHeadToNext() {
+        this.head = this.head.next;
+        if (this.head) {
+            this.head.prev = null;
+        } else {
+            this.tail = null;
+        }
+    }
+}
+
+export { FastLRUCache };

package/lib/pathwayTools.js

@@ -1,5 +1,5 @@
 // pathwayTools.js
-import { encode , decode } from 'gpt-3-encoder';
+import { encode, decode } from '../lib/encodeCache.js';
 import { config } from '../config.js';
 
 // callPathway - call a pathway from another pathway

package/lib/requestExecutor.js

@@ -57,9 +57,10 @@ const createLimiter = (endpoint, name, index) => {
 
     endpoint.limiter.on('failed', (error, info) => {
         if (error.name === 'CanceledError') {
-            logger.debug(`Request cancelled for ${cortexId}-${name}-${index}: Id: ${info.options.id || 'none'}`);
+            logger.debug(`Limiter request cancelled for ${cortexId}-${name}-${index}: Id: ${info.options.id || 'none'}`);
+            endpoint.monitor.incrementErrorCount();
         } else {
-            logger.error(`Request failed for ${cortexId}-${name}-${index}: Id: ${info.options.id || 'none'}: ${error}`);
+            logger.error(`Limiter request failed for ${cortexId}-${name}-${index}: Id: ${info.options.id || 'none'}: ${error?.message || error}`);
         }
     });
 
@@ -154,6 +155,7 @@ if (config.get('enableCache')) {
     }); 
 }
 
+//log statistics about active endpoints
 setInterval(() => {
   // Iterate over each model
   for (const [name, model] of Object.entries(modelEndpoints)) {
@@ -179,100 +181,106 @@ setInterval(() => {
         endpointIndex++;
     });
   }
-}, 10000); // Log rates every 10 seconds (10000 ms).
+}, 30000); // Log rates every 30 seconds
 
-const postWithMonitor = async (endpoint, url, data, axiosConfigObj) => {
-    return cortexAxios.post(url, data, axiosConfigObj);
+const requestWithMonitor = async (endpoint, url, data, axiosConfigObj) => {
+    const callId = endpoint?.monitor?.startCall();
+    let response;
+    try {
+        if (axiosConfigObj?.method == 'GET'){
+            response = await cortexAxios.get(url, axiosConfigObj);
+        } else {
+            response = await cortexAxios.post(url, data, axiosConfigObj);
+        }
+    } catch (error) {
+        // throw new error with duration as part of the error data
+        throw { ...error, duration: endpoint?.monitor?.incrementErrorCount(callId, error?.response?.status || null) };
+    }
+    let duration;
+    if (response.status >= 200 && response.status < 300) {
+        duration = endpoint?.monitor?.endCall(callId);
+    } else {
+        duration = endpoint?.monitor?.incrementErrorCount(callId, response.status);
+    }
+
+    return { response, duration };
 }
 
 const MAX_RETRY = 10; // retries for error handling
 const MAX_DUPLICATE_REQUESTS = 3; // duplicate requests to manage latency spikes
 const DUPLICATE_REQUEST_AFTER = 10; // 10 seconds
 
-const postRequest = async (cortexRequest) => {
+const getDuplicateRequestDelay = (index, duplicateRequestAfter) => {
+    const duplicateRequestTime = duplicateRequestAfter * Math.pow(2, index) - duplicateRequestAfter;
+    const jitter = duplicateRequestTime * 0.2 * Math.random();
+    const duplicateRequestTimeout = Math.max(0, duplicateRequestTime + jitter);
+    return duplicateRequestTimeout;
+}
+
+const makeRequest = async (cortexRequest) => {
     let promises = [];
+    // retry certain errors up to MAX_RETRY times
     for (let i = 0; i < MAX_RETRY; i++) {
-        const { url, data, params, headers, cache, selectedEndpoint, requestId, pathway, model, stream} = cortexRequest;
+        const { url, data, params, headers, cache, selectedEndpoint, requestId, pathway, model, stream, method} = cortexRequest;
         const enableDuplicateRequests = pathway?.enableDuplicateRequests !== undefined ? pathway.enableDuplicateRequests : config.get('enableDuplicateRequests');
-        let maxDuplicateRequests = enableDuplicateRequests ? MAX_DUPLICATE_REQUESTS : 1;
-        let duplicateRequestAfter = (pathway?.duplicateRequestAfter || DUPLICATE_REQUEST_AFTER) * 1000;
-
-        if (enableDuplicateRequests) {
-            //logger.info(`>>> [${requestId}] Duplicate requests enabled after ${duplicateRequestAfter / 1000} seconds`);
-        }
+        const maxDuplicateRequests = enableDuplicateRequests ? MAX_DUPLICATE_REQUESTS : 1;
+        const duplicateRequestAfter = (pathway?.duplicateRequestAfter || DUPLICATE_REQUEST_AFTER) * 1000;
 
-        const axiosConfigObj = { params, headers, cache };
+        const axiosConfigObj = { params, headers, cache, method };
         const streamRequested = (stream || params?.stream || data?.stream);
+        // if we're using streaming, duplicate requests are
+        // not supported, so we just push one promise into the array
         if (streamRequested && model.supportsStreaming) {
             axiosConfigObj.responseType = 'stream';
-            promises.push(selectedEndpoint.limiter.schedule({expiration: pathway.timeout * 1000 + 1000, id: `${requestId}_${uuidv4()}`},() => postWithMonitor(selectedEndpoint, url, data, axiosConfigObj)));
+            promises.push(selectedEndpoint.limiter.schedule({expiration: pathway.timeout * 1000 + 1000, id: `${requestId}_${uuidv4()}`},() => requestWithMonitor(selectedEndpoint, url, data, axiosConfigObj)));
         } else {
             if (streamRequested) {
                 logger.info(`>>> [${requestId}] ${model} does not support streaming - sending non-streaming request`);
                 axiosConfigObj.params.stream = false;
                 data.stream = false;
             }
+            // if we're not streaming, we push at least one promise
+            // into the array, but if we're supporting duplicate
+            // requests we push one for each potential duplicate,
+            // heading to a new endpoint (if available) and
+            // staggered by a jittered amount of time
             const controllers = Array.from({ length: maxDuplicateRequests }, () => new AbortController());
             promises = controllers.map((controller, index) =>
                 new Promise((resolve, reject) => {
-                    const duplicateRequestTime = duplicateRequestAfter * Math.pow(2, index) - duplicateRequestAfter;
-                    const jitter = duplicateRequestTime * 0.2 * Math.random();
-                    const duplicateRequestTimeout = Math.max(0, duplicateRequestTime + jitter);
                     setTimeout(async () => {
                         try {
+                            if (index > 0) {
+                                cortexRequest.selectNewEndpoint();
+                            }
+                            const { url, data, params, headers, cache, selectedEndpoint, requestId, pathway, model } = cortexRequest;
                             const endpointName = selectedEndpoint.name || model;
                             if (!selectedEndpoint.limiter) {
                                 throw new Error(`No limiter for endpoint ${endpointName}!`);
                             }
-                            const axiosConfigObj = { params, headers, cache };
+                            const axiosConfigObj = { params, headers, cache, method };
 
                             let response = null;
+                            let duration = null;
 
                             if (!controller.signal?.aborted) {
 
                                 axiosConfigObj.signal = controller.signal;
                                 axiosConfigObj.headers['X-Cortex-Request-Index'] = index;
 
-                                if (index === 0) {
-                                    //logger.info(`>>> [${requestId}] sending request to ${endpointName} API ${axiosConfigObj.responseType === 'stream' ? 'with streaming' : ''}`);
-                                } else {
-                                    if (model.supportsStreaming) {
-                                        axiosConfigObj.responseType = 'stream';
-                                        axiosConfigObj.cache = false;
-                                    }
-                                    const logMessage = `>>> [${requestId}] taking too long - sending duplicate request ${index} to ${endpointName} API ${axiosConfigObj.responseType === 'stream' ? 'with streaming' : ''}`;
+                                if (index > 0) {
+                                    const logMessage = `>>> [${requestId}] taking too long - sending duplicate request ${index} to ${endpointName} API`;
                                     const header = '>'.repeat(logMessage.length);
                                     logger.info(`\n${header}\n${logMessage}`);
                                 }
 
-                                response = await selectedEndpoint.limiter.schedule({expiration: pathway.timeout * 1000 + 1000, id: `${requestId}_${uuidv4()}`}, () => postWithMonitor(selectedEndpoint, url, data, axiosConfigObj));
+                                ({ response, duration } = await selectedEndpoint.limiter.schedule({expiration: pathway.timeout * 1000 + 1000, id: `${requestId}_${uuidv4()}`}, () => requestWithMonitor(selectedEndpoint, url, data, axiosConfigObj)));
 
                                 if (!controller.signal?.aborted) {
-
                                     logger.debug(`<<< [${requestId}] received response for request ${index}`);
-
-                                    if (axiosConfigObj.responseType === 'stream') {
-                                        // Buffering and collecting the stream data
-                                        logger.info(`<<< [${requestId}] buffering streaming response for request ${index}`);
-                                        response = await new Promise((resolve, reject) => {
-                                            let responseData = '';
-                                            response.data.on('data', (chunk) => {
-                                                responseData += chunk;
-                                                logger.debug(`<<< [${requestId}] received chunk for request ${index}`);
-                                            });
-                                            response.data.on('end', () => {
-                                                response.data = JSON.parse(responseData);
-                                                resolve(response);
-                                            });
-                                            response.data.on('error', (error) => {
-                                                reject(error);
-                                            });
-                                        });
-                                    }
                                 }
                             }
 
-                            resolve(response);
+                            resolve({ response, duration });
 
                         } catch (error) {
                             if (error.name === 'AbortError' || error.name === 'CanceledError') {
@@ -285,45 +293,48 @@ const postRequest = async (cortexRequest) => {
                         } finally {
                             controllers.forEach(controller => controller.abort());
                         }
-                    }, duplicateRequestTimeout);
+                    }, getDuplicateRequestDelay(index, duplicateRequestAfter));
                 })
             );
         }
 
+        // no requests have been made yet, but the promises array
+        // is full, so now we execute them in parallel
         try {
-            const response = await Promise.race(promises);
+            const { response, duration } = await Promise.race(promises);
 
             // if response status is 2xx
             if (response.status >= 200 && response.status < 300) {
-                return response;
+                return { response, duration };
             } else {
                 throw new Error(`Received error response: ${response.status}`);
             }
         } catch (error) {
-            if (error.response) {
-                selectedEndpoint.monitor.incrementErrorCount();
-                const status = error.response.status;
-
-                if (status === 429) {
-                    selectedEndpoint.monitor.incrementError429Count();
-                }
- 
+            const { response, duration } = error;
+            if (response) {
+                const status = response.status;
+                // if there is only one endpoint, only retry select error codes
                 if (cortexRequest.model.endpoints.length === 1) {
-                    if (status !== 429) {
-                        return error.response;
+                    if (status !== 429 &&
+                        status !== 408 &&
+                        status !== 502 &&
+                        status !== 503 &&
+                        status !== 504) {
+                        return { response, duration };
                     }
                 } else {
-                    // if there are multiple endpoints, retry everything
+                    // if there are multiple endpoints, retry everything as it
+                    // could be going to a different host
                     cortexRequest.selectNewEndpoint();
                 }
 
-                logger.info(`>>> [${requestId}] retrying request due to ${status} response. Retry count: ${i + 1}`);
+                logger.info(`>>> [${requestId}] retrying request (${duration}ms) due to ${status} response. Retry count: ${i + 1}`);
                 if (i < MAX_RETRY - 1) {
                     const backoffTime = 200 * Math.pow(2, i);
                     const jitter = backoffTime * 0.2 * Math.random();
                     await new Promise(r => setTimeout(r, backoffTime + jitter));
                 } else {
-                    return error.response;
+                    return { response, duration };
                 }
             } else {
                 throw error;
@@ -334,10 +345,7 @@ const postRequest = async (cortexRequest) => {
 
 const executeRequest = async (cortexRequest) => {
     try {
-        const endpoint = cortexRequest.selectedEndpoint;
-        const callId = endpoint?.monitor?.startCall();
-        const response = await postRequest(cortexRequest);
-        endpoint?.monitor?.endCall(callId);
+        const { response, duration } = await makeRequest(cortexRequest);
         const requestId = cortexRequest.requestId;
         const { error, data, cached } = response;
         if (cached) {
@@ -347,8 +355,7 @@ const executeRequest = async (cortexRequest) => {
             const lastError = error[error.length - 1];
             return { error: lastError.toJSON() ?? lastError ?? error };
         }
-        //logger.info(`<<< [${requestId}] response: ${data.choices[0].delta || data.choices[0]}`)
-        return data;
+        return { data, duration };
     } catch (error) {
         logger.error(`Error in request: ${error.message || error}`);
         return { error: error };

package/lib/requestMonitor.js

@@ -1,5 +1,4 @@
 import { v4 as uuidv4 } from 'uuid';
-// eslint-disable-next-line import/no-extraneous-dependencies
 import { Deque } from '@datastructures-js/deque';
 
 class RequestMonitor {
@@ -20,6 +19,15 @@ class RequestMonitor {
     return this.healthy;
   }
 
+  removeOldCallStarts() {
+    const currentTime = new Date();
+    for (const [callId, startTime] of this.callStartTimes) {
+      if (currentTime - startTime > this.ageOutTime) {
+        this.callStartTimes.delete(callId);
+      }
+    }
+  }
+
   removeOldCallStats(dq, timeProperty) {
     const currentTime = new Date();
     while (!dq.isEmpty() && currentTime - (timeProperty ? dq.front()[timeProperty] : dq.front())  > this.ageOutTime) {
@@ -28,6 +36,7 @@ class RequestMonitor {
   }
   
   maintain() {
+    this.removeOldCallStarts();
     this.removeOldCallStats(this.callCount);
     if (this.callCount.size() === 0) {
       this.peakCallRate = 0;
@@ -36,7 +45,7 @@ class RequestMonitor {
     this.removeOldCallStats(this.error429Count);
     this.removeOldCallStats(this.errorCount);
   
-    if (this.getErrorRate() > 0.3) {
+    if (this.getErrorRate() > 0.1) {
       this.healthy = false;
     } else {
       this.healthy = true;
@@ -55,10 +64,11 @@ class RequestMonitor {
   endCall(callId) {
     const endTime = new Date();
     const startTime = this.callStartTimes.get(callId);
+    let callDuration = null;
 
     if (startTime) {
+      callDuration = (endTime - startTime);
       this.callStartTimes.delete(callId);
-      const callDuration = endTime - startTime;
       this.callDurations.pushBack({endTime, callDuration});
 
       // Keep the callDurations length to 5
@@ -73,6 +83,7 @@ class RequestMonitor {
     }
     
     this.maintain();
+    return callDuration;
   }
 
   getAverageCallDuration() {
@@ -84,14 +95,13 @@ class RequestMonitor {
     return sum / this.callDurations.size(); 
   }
 
-  incrementError429Count() {
-    this.error429Count.pushBack(new Date());
-    this.maintain();
-  }
-
-  incrementErrorCount() {
+  incrementErrorCount(callId, status) {
     this.errorCount.pushBack(new Date());
+    if (status === 429) {
+      this.error429Count.pushBack(new Date());
+    }
     this.maintain();
+    return callId ? this.endCall(callId) : null;
   }
 
   getCallRate() {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aj-archipelago/cortex",
-  "version": "1.1.5",
+  "version": "1.1.7",
   "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": {
@@ -52,7 +52,7 @@
     "handlebars": "^4.7.7",
     "ioredis": "^5.3.1",
     "keyv": "^4.5.2",
-    "langchain": "^0.0.47",
+    "langchain": "^0.1.28",
     "mime-types": "^2.1.35",
     "subsrt": "^1.1.1",
     "uuid": "^9.0.0",
@@ -60,6 +60,7 @@
     "ws": "^8.12.0"
   },
   "devDependencies": {
+    "@faker-js/faker": "^8.4.1",
     "ava": "^5.2.0",
     "dotenv": "^16.0.3",
     "eslint": "^8.38.0",

package/pathways/basePathway.js

@@ -14,19 +14,21 @@ export default {
     typeDef,
     rootResolver,
     resolver,
-    inputFormat: 'text', // text or html - changes the behavior of the input chunking
+    inputFormat: 'text', // string - 'text' or 'html' - changes the behavior of the input chunking
     useInputChunking: true, // true or false - enables input to be split into multiple chunks to meet context window size
     useParallelChunkProcessing: false, // true or false - enables parallel processing of chunks
+    joinChunksWith: '\n\n', // string - the string to join result chunks with when useInputChunking is 'true'
     useInputSummarization: false, // true or false - instead of chunking, summarize the input and act on the summary    
     truncateFromFront: false, // true or false - if true, truncate from the front of the input instead of the back
     timeout: 120, // seconds, cancels the pathway after this many seconds
+    enableDuplicateRequests: true, // true or false - if true, duplicate requests are sent if the request is not completed after duplicateRequestAfter seconds
     duplicateRequestAfter: 10, // seconds, if the request is not completed after this many seconds, a backup request is sent
     // override the default execution of the pathway
-    // callback signature: excuteOverride({args: object, runAllPrompts: function})
+    // callback signature: executeOverride({args: object, runAllPrompts: function})
     // args: the input arguments to the pathway
     // runAllPrompts: a function that runs all prompts in the pathway and returns the result
     executePathway: undefined,
     // Set the temperature to 0 to favor more deterministic output when generating entity extraction.
-    temperature: undefined,
+    temperature: 0.9,
 };
 

package/pathways/bing.js

@@ -0,0 +1,12 @@
+// bing.js
+// Web search tool
+
+export default {
+    inputParameters: {
+        text: ``,
+    },
+    timeout: 400,
+    enableDuplicateRequests: false,
+    model: 'azure-bing',
+};
+

package/pathways/index.js

@@ -19,8 +19,10 @@ import transcribe from './transcribe.js';
 import translate from './translate.js';
 import embeddings from './embeddings.js';
 import vision from './vision.js';
+import bing from './bing.js';
 
 export {
+    bing,
     edit,
     chat,
     bias,

package/server/chunker.js

@@ -1,4 +1,4 @@
-import { encode, decode } from 'gpt-3-encoder';
+import { encode, decode } from '../lib/encodeCache.js';
 import cheerio from 'cheerio';
 
 const getLastNToken = (text, maxTokenLen) => { 

package/server/graphql.js

@@ -131,7 +131,7 @@ const build = async (config) => {
 
     const app = express();
 
-    app.use(express.json({ limit: '50mb' }));
+    app.use(express.json({ limit: '200mb' }));
 
     const httpServer = http.createServer(app);
 

package/server/modelExecutor.js

@@ -19,6 +19,7 @@ import OpenAIDallE3Plugin from './plugins/openAiDallE3Plugin.js';
 import OpenAIVisionPlugin from './plugins/openAiVisionPlugin.js';
 import GeminiChatPlugin from './plugins/geminiChatPlugin.js';
 import GeminiVisionPlugin from './plugins/geminiVisionPlugin.js';
+import AzureBingPlugin from './plugins/azureBingPlugin.js';
 
 class ModelExecutor {
     constructor(pathway, model) {
@@ -80,6 +81,9 @@ class ModelExecutor {
             case 'GEMINI-VISION':
                 plugin = new GeminiVisionPlugin(pathway, model);
                 break;
+            case 'AZURE-BING':
+                plugin = new AzureBingPlugin(pathway, model);
+                break;
             default:
                 throw new Error(`Unsupported model type: ${model.type}`);
         }