@aj-archipelago/cortex 1.3.35 → 1.3.36

package/server/plugins/palmChatPlugin.js

@@ -1,233 +0,0 @@
-// palmChatPlugin.js
-import ModelPlugin from './modelPlugin.js';
-import HandleBars from '../../lib/handleBars.js';
-import logger from '../../lib/logger.js';
-
-class PalmChatPlugin extends ModelPlugin {
-    constructor(pathway, model) {
-        super(pathway, model);
-    }
-
-    // Convert to PaLM messages array format if necessary
-    convertMessagesToPalm(messages) {
-        let context = '';
-        let modifiedMessages = [];
-        let lastAuthor = '';
-    
-        // remove any empty messages
-        messages = messages.filter(message => message.content);
-
-        messages.forEach(message => {
-            const { role, author, content } = message;
-    
-            // Extract system messages into the context string
-            if (role === 'system') {
-                context += (context.length > 0 ? '\n' : '') + content;
-                return;
-            }
-     
-            // Aggregate consecutive author messages, appending the content
-            if ((role === lastAuthor || author === lastAuthor) && modifiedMessages.length > 0) {
-                modifiedMessages[modifiedMessages.length - 1].content += '\n' + content;
-            }
-            // Only push messages with role 'user' or 'assistant' or existing author messages
-            else if (role === 'user' || role === 'assistant' || author) {
-                modifiedMessages.push({
-                    author: author || role,
-                    content,
-                });
-                lastAuthor = author || role;
-            }
-        });
-    
-        return {
-            modifiedMessages,
-            context,
-        };
-    }
-
-    // Handlebars compiler for context (PaLM chat specific)
-    getCompiledContext(text, parameters, context) {
-        const combinedParameters = { ...this.promptParameters, ...parameters };
-        return context ? HandleBars.compile(context)({ ...combinedParameters, text}) : '';
-    }
-
-    // Handlebars compiler for examples (PaLM chat specific)
-    getCompiledExamples(text, parameters, examples = []) {
-        const combinedParameters = { ...this.promptParameters, ...parameters };
-
-        const compileContent = (content) => {
-            const compile = HandleBars.compile(content);
-            return compile({ ...combinedParameters, text });
-        };
-
-        const processExample = (example, key) => {
-            if (example[key]?.content) {
-                return { ...example[key], content: compileContent(example[key].content) };
-            }
-            return { ...example[key] };
-        };
-
-        return examples.map((example) => ({
-            input: example.input ? processExample(example, 'input') : undefined,
-            output: example.output ? processExample(example, 'output') : undefined,
-        }));
-    }
-
-    // Set up parameters specific to the PaLM Chat API
-    getRequestParameters(text, parameters, prompt) {
-        const { modelPromptText, modelPromptMessages, tokenLength } = this.getCompiledPrompt(text, parameters, prompt);
-    
-        // Define the model's max token length
-        const modelTargetTokenLength = this.getModelMaxPromptTokens();
-    
-        const palmMessages = this.convertMessagesToPalm(modelPromptMessages || [{ "author": "user", "content": modelPromptText }]);
-        
-        let requestMessages = palmMessages.modifiedMessages;
-
-        // Check if the token length exceeds the model's max token length
-        if (tokenLength > modelTargetTokenLength) {
-            // Remove older messages until the token length is within the model's limit
-            requestMessages = this.truncateMessagesToTargetLength(requestMessages, modelTargetTokenLength);
-        }
-
-        const context = this.getCompiledContext(text, parameters, prompt.context || palmMessages.context || '');
-        const examples = this.getCompiledExamples(text, parameters, prompt.examples || []);
-        
-        const max_tokens = this.getModelMaxReturnTokens();
-        
-        if (max_tokens < 0) {
-            throw new Error(`Prompt is too long to successfully call the model at ${tokenLength} tokens.  The model will not be called.`);
-        }
-    
-        // Ensure there are an odd number of messages for turn taking
-        if (requestMessages.length % 2 === 0) {
-            requestMessages = requestMessages.slice(1);
-        }
-
-        const requestParameters = {
-            instances: [{
-                context: context,
-                examples: examples,
-                messages: requestMessages,
-            }],
-            parameters: {
-                temperature: this.temperature ?? 0.7,
-                maxOutputTokens: max_tokens,
-                topP: parameters.topP ?? 0.95,
-                topK: parameters.topK ?? 40,
-            }
-        };
-    
-        return requestParameters;
-    }
-
-    // Get the safetyAttributes from the PaLM Chat API response data
-    getSafetyAttributes(data) {
-        const { predictions } = data;
-        if (!predictions || !predictions.length) {
-            return null;
-        }
-
-        // if we got a predictions array back with more than one prediction, return the safetyAttributes of the first prediction
-        if (predictions.length > 1) {
-            return predictions[0].safetyAttributes ?? null;
-        }
-
-        // otherwise, return the safetyAttributes of the content of the first prediction
-        return predictions[0].safetyAttributes ?? null;
-    }
-
-    // Execute the request to the PaLM Chat API
-    async execute(text, parameters, prompt, cortexRequest) {
-        const requestParameters = this.getRequestParameters(text, parameters, prompt);
-
-        cortexRequest.data = { ...(cortexRequest.data || {}), ...requestParameters };
-        cortexRequest.params = {}; // query params
-
-        const gcpAuthTokenHelper = this.config.get('gcpAuthTokenHelper');
-        const authToken = await gcpAuthTokenHelper.getAccessToken();
-        cortexRequest.auth.Authorization = `Bearer ${authToken}`;
-
-        return this.executeRequest(cortexRequest);
-    }
-
-    // Parse the response from the PaLM Chat API
-    parseResponse(data) {
-        const { predictions } = data;
-        if (!predictions || !predictions.length) {
-            return data;
-        }
-    
-        // Get the candidates array from the first prediction
-        const { candidates } = predictions[0];
-
-        // if it was blocked, return the blocked message
-        if (predictions[0].safetyAttributes?.blocked) {
-            return 'The response is blocked because the input or response potentially violates Google policies. Try rephrasing the prompt or adjusting the parameter settings. Currently, only English is supported.';
-        }
-
-        if (!candidates || !candidates.length) {
-            return null;
-        }
-    
-        // If we got a candidates array back with more than one candidate, return the whole array
-        if (candidates.length > 1) {
-            return candidates;
-        }
-    
-        // Otherwise, return the content of the first candidate
-        const messageResult = candidates[0].content && candidates[0].content.trim();
-        return messageResult ?? null;
-    }
-
-    // Override the logging function to display the messages and responses
-    logRequestData(data, responseData, prompt) {
-        const instances = data && data.instances;
-        const messages = instances && instances[0] && instances[0].messages;
-        const { context, examples } = instances && instances [0] || {};
-    
-        if (context) {
-            const { length, units } = this.getLength(context);
-            logger.info(`[chat request contains context information of length ${length} ${units}]`)
-            logger.verbose(`context: ${context}`);
-        }
-
-        if (examples && examples.length) {
-            logger.info(`[chat request contains ${examples.length} examples]`);
-            examples.forEach((example, index) => {
-                logger.verbose(`example ${index + 1}: input: "${example.input.content}", output: "${example.output.content}"`);
-            });
-        }
-        
-        if (messages && messages.length > 1) {
-            logger.info(`[chat request contains ${messages.length} messages]`);
-            messages.forEach((message, index) => {
-                const words = message.content.split(" ");
-                const { length, units } = this.getLength(message.content);
-                const preview = words.length < 41 ? message.content : words.slice(0, 20).join(" ") + " ... " + words.slice(-20).join(" ");
-    
-                logger.verbose(`message ${index + 1}: author: ${message.author}, ${units}: ${length}, content: "${preview}"`);
-            });
-        } else if (messages && messages.length === 1) {
-            logger.verbose(`${messages[0].content}`);
-        }
-
-        const safetyAttributes = this.getSafetyAttributes(responseData);
-
-        const responseText = this.parseResponse(responseData);
-        const { length, units } = this.getLength(responseText);
-        logger.info(`[response received containing ${length} ${units}]`);
-        logger.verbose(`${responseText}`);
-
-        if (safetyAttributes) {
-            logger.warn(`[response contains safety attributes: ${JSON.stringify(safetyAttributes, null, 2)}]`);
-        }
-    
-        if (prompt && prompt.debugInfo) {
-            prompt.debugInfo += `\n${JSON.stringify(data)}`;
-        }
-    }
-}
-
-export default PalmChatPlugin;

package/server/plugins/palmCodeCompletionPlugin.js

@@ -1,45 +0,0 @@
-// palmCodeCompletionPlugin.js
-
-import PalmCompletionPlugin from './palmCompletionPlugin.js';
-
-// PalmCodeCompletionPlugin class for handling requests and responses to the PaLM API Code Completion API
-class PalmCodeCompletionPlugin extends PalmCompletionPlugin {
-    constructor(pathway, model) {
-        super(pathway, model);
-    }
-
-    // Set up parameters specific to the PaLM API Code Completion API
-    getRequestParameters(text, parameters, prompt, pathwayResolver) {
-        const { modelPromptText, tokenLength } = this.getCompiledPrompt(text, parameters, prompt);
-        // Define the model's max token length
-        const modelTargetTokenLength = this.getModelMaxPromptTokens();
-    
-        const truncatedPrompt = this.truncatePromptIfNecessary(modelPromptText, tokenLength, this.getModelMaxTokenLength(), modelTargetTokenLength, pathwayResolver);
-    
-        const max_tokens = this.getModelMaxReturnTokens();
-    
-        if (max_tokens < 0) {
-            throw new Error(`Prompt is too long to successfully call the model at ${tokenLength} tokens.  The model will not be called.`);
-        }
-
-        if (!truncatedPrompt) {
-            throw new Error(`Prompt is empty.  The model will not be called.`);
-        }
-    
-        const requestParameters = {
-            instances: [
-                { prefix: truncatedPrompt }
-            ],
-            parameters: {
-                temperature: this.temperature ?? 0.7,
-                maxOutputTokens: max_tokens,
-                topP: parameters.topP ?? 0.95,
-                topK: parameters.topK ?? 40,
-            }
-        };
-    
-        return requestParameters;
-    }
-}
-
-export default PalmCodeCompletionPlugin;

package/server/plugins/palmCompletionPlugin.js

@@ -1,135 +0,0 @@
-// palmCompletionPlugin.js
-
-import ModelPlugin from './modelPlugin.js';
-import logger from '../../lib/logger.js';
-
-// PalmCompletionPlugin class for handling requests and responses to the PaLM API Text Completion API
-class PalmCompletionPlugin extends ModelPlugin {
-    constructor(pathway, model) {
-        super(pathway, model);
-    }
-
-    truncatePromptIfNecessary (text, textTokenCount, modelMaxTokenCount, targetTextTokenCount, pathwayResolver) {
-        const maxAllowedTokens = textTokenCount + ((modelMaxTokenCount - targetTextTokenCount) * 0.5);
-    
-        if (textTokenCount > maxAllowedTokens) {
-            pathwayResolver.logWarning(`Prompt is too long at ${textTokenCount} tokens (this target token length for this pathway is ${targetTextTokenCount} tokens because the response is expected to take up the rest of the model's max tokens (${modelMaxTokenCount}). Prompt will be truncated.`);
-            return pathwayResolver.truncate(text, maxAllowedTokens);
-        }
-        return text;
-    }
-    // Set up parameters specific to the PaLM API Text Completion API
-    getRequestParameters(text, parameters, prompt, pathwayResolver) {
-        const { modelPromptText, tokenLength } = this.getCompiledPrompt(text, parameters, prompt);
-        
-        // Define the model's max token length
-        const modelTargetTokenLength = this.getModelMaxPromptTokens();
-    
-        const truncatedPrompt = this.truncatePromptIfNecessary(modelPromptText, tokenLength, this.getModelMaxTokenLength(), modelTargetTokenLength, pathwayResolver);
-    
-        const max_tokens = this.getModelMaxReturnTokens();
-    
-        if (max_tokens < 0) {
-            throw new Error(`Prompt is too long to successfully call the model at ${tokenLength} tokens.  The model will not be called.`);
-        }
-
-        if (!truncatedPrompt) {
-            throw new Error(`Prompt is empty.  The model will not be called.`);
-        }
-    
-        const requestParameters = {
-            instances: [
-                { prompt: truncatedPrompt }
-            ],
-            parameters: {
-                temperature: this.temperature ?? 0.7,
-                maxOutputTokens: max_tokens,
-                topP: parameters.topP ?? 0.95,
-                topK: parameters.topK ?? 40,
-            }
-        };
-    
-        return requestParameters;
-    }
-
-    // Execute the request to the PaLM API Text Completion API
-    async execute(text, parameters, prompt, cortexRequest) {
-        const requestParameters = this.getRequestParameters(text, parameters, prompt, cortexRequest.pathwayResolver);
-
-        cortexRequest.data = { ...(cortexRequest.data || {}), ...requestParameters };
-        cortexRequest.params = {}; // query params
-
-        const gcpAuthTokenHelper = this.config.get('gcpAuthTokenHelper');
-        const authToken = await gcpAuthTokenHelper.getAccessToken();
-        cortexRequest.auth.Authorization = `Bearer ${authToken}`;
-
-        return this.executeRequest(cortexRequest);
-    }
-
-    // Parse the response from the PaLM API Text Completion API
-    parseResponse(data) {
-        const { predictions } = data;
-        if (!predictions || !predictions.length) {
-            return data;
-        }
-
-        // if we got a predictions array back with more than one prediction, return the whole array
-        if (predictions.length > 1) {
-            return predictions;
-        }
-
-        // otherwise, return the content of the first prediction
-        // if it was blocked, return the blocked message
-        if (predictions[0].safetyAttributes?.blocked) {
-            return 'The response is blocked because the input or response potentially violates Google policies. Try rephrasing the prompt or adjusting the parameter settings. Currently, only English is supported.';
-        }
-
-        const contentResult = predictions[0].content && predictions[0].content.trim();
-        return contentResult ?? null;
-    }
-
-    // Get the safetyAttributes from the PaLM API Text Completion API response data
-    getSafetyAttributes(data) {
-        const { predictions } = data;
-        if (!predictions || !predictions.length) {
-            return null;
-        }
-
-        // if we got a predictions array back with more than one prediction, return the safetyAttributes of the first prediction
-        if (predictions.length > 1) {
-            return predictions[0].safetyAttributes ?? null;
-        }
-
-        // otherwise, return the safetyAttributes of the content of the first prediction
-        return predictions[0].safetyAttributes ?? null;
-    }
-
-    // Override the logging function to log the prompt and response
-    logRequestData(data, responseData, prompt) {
-        const safetyAttributes = this.getSafetyAttributes(responseData);
-
-        const instances = data && data.instances;
-        const modelInput = instances && instances[0] && instances[0].prompt;
-
-        if (modelInput) {
-            const { length, units } = this.getLength(modelInput);
-            logger.info(`[request sent containing ${length} ${units}]`);
-            logger.verbose(`${modelInput}`);
-        }
-
-        const responseText = this.parseResponse(responseData);
-        const { length, units } = this.getLength(responseText);
-        logger.info(`[response received containing ${length} ${units}]`);
-        logger.verbose(`${responseText}`);
-
-        if (safetyAttributes) {
-            logger.warn(`[response contains safety attributes: ${JSON.stringify(safetyAttributes, null, 2)}]`);
-        }
-
-        if (prompt && prompt.debugInfo) {
-            prompt.debugInfo += `\n${JSON.stringify(data)}`;
-        }
-    }
-}
-
-export default PalmCompletionPlugin;

package/tests/palmChatPlugin.test.js

@@ -1,219 +0,0 @@
-// test_palmChatPlugin.js
-import test from 'ava';
-import PalmChatPlugin from '../server/plugins/palmChatPlugin.js';
-import { mockPathwayResolverMessages } from './mocks.js';
-
-const { config, pathway, modelName, model } = mockPathwayResolverMessages;
-
-test.beforeEach((t) => {
-  const palmChatPlugin = new PalmChatPlugin(pathway, model);
-  t.context = { palmChatPlugin };
-});
-
-test('convertMessagesToPalm', (t) => {
-  const { palmChatPlugin } = t.context;
-  const messages = [
-    { role: 'system', content: 'System Message' },
-    { role: 'user', content: 'User Message' },
-    { role: 'user', content: 'User Message 2'},
-  ];
-
-  const expectedResult = {
-    modifiedMessages: [
-      { author: 'user', content: 'User Message\nUser Message 2' },
-    ],
-    context: 'System Message',
-  };
-
-  t.deepEqual(palmChatPlugin.convertMessagesToPalm(messages), expectedResult);
-});
-
-test('convertMessagesToPalm - already PaLM format', (t) => {
-    const { palmChatPlugin } = t.context;
-    const messages = [
-      { author: 'user', content: 'User Message' },
-      { author: 'user', content: 'User Message 2'},
-    ];
-  
-    const expectedResult = {
-      modifiedMessages: [
-        { author: 'user', content: 'User Message\nUser Message 2' },
-      ],
-      context: '',
-    };
-  
-    t.deepEqual(palmChatPlugin.convertMessagesToPalm(messages), expectedResult);
-  });
-
-  test('convertMessagesToPalm - empty string roles', (t) => {
-    const { palmChatPlugin } = t.context;
-    const messages = [
-      { role: '', content: 'Empty role message' },
-      { role: 'user', content: 'User Message' },
-    ];
-  
-    const expectedResult = {
-      modifiedMessages: [
-        { author: 'user', content: 'User Message' },
-      ],
-      context: '',
-    };
-  
-    t.deepEqual(palmChatPlugin.convertMessagesToPalm(messages), expectedResult);
-  });
-  
-  test('convertMessagesToPalm - consecutive system messages', (t) => {
-    const { palmChatPlugin } = t.context;
-    const messages = [
-      { role: 'system', content: 'System Message 1' },
-      { role: 'system', content: 'System Message 2' },
-      { role: 'user', content: 'User Message' },
-    ];
-  
-    const expectedResult = {
-      modifiedMessages: [
-        { author: 'user', content: 'User Message' },
-      ],
-      context: 'System Message 1\nSystem Message 2',
-    };
-  
-    t.deepEqual(palmChatPlugin.convertMessagesToPalm(messages), expectedResult);
-  });
-  
-  test('convertMessagesToPalm - multiple authors', (t) => {
-    const { palmChatPlugin } = t.context;
-    const messages = [
-      { role: 'system', content: 'System Message' },
-      { author: 'user1', content: 'User1 Message' },
-      { author: 'user1', content: 'User1 Message 2' },
-      { author: 'user2', content: 'User2 Message' },
-      { author: 'assistant', content: 'Assistant Message' },
-    ];
-  
-    const expectedResult = {
-      modifiedMessages: [
-        { author: 'user1', content: 'User1 Message\nUser1 Message 2' },
-        { author: 'user2', content: 'User2 Message' },
-        { author: 'assistant', content: 'Assistant Message' },
-      ],
-      context: 'System Message',
-    };
-  
-    t.deepEqual(palmChatPlugin.convertMessagesToPalm(messages), expectedResult);
-  });
-  
-  test('convertMessagesToPalm - no messages', (t) => {
-    const { palmChatPlugin } = t.context;
-    const messages = [];
-  
-    const expectedResult = {
-      modifiedMessages: [],
-      context: '',
-    };
-  
-    t.deepEqual(palmChatPlugin.convertMessagesToPalm(messages), expectedResult);
-  });
-  
-  test('convertMessagesToPalm - only system messages', (t) => {
-    const { palmChatPlugin } = t.context;
-    const messages = [
-      { role: 'system', content: 'System Message 1' },
-      { role: 'system', content: 'System Message 2' },
-    ];
-  
-    const expectedResult = {
-      modifiedMessages: [],
-      context: 'System Message 1\nSystem Message 2',
-    };
-  
-    t.deepEqual(palmChatPlugin.convertMessagesToPalm(messages), expectedResult);
-  });
-
-test('getCompiledContext', (t) => {
-  const { palmChatPlugin } = t.context;
-  const text = 'Hello';
-  const parameters = { name: 'John' };
-  const context = '{{text}} from {{name}}';
-
-  const expectedResult = 'Hello from John';
-
-  t.is(palmChatPlugin.getCompiledContext(text, parameters, context), expectedResult);
-});
-
-test('getCompiledExamples', (t) => {
-  const { palmChatPlugin } = t.context;
-  const text = 'Greetings';
-  const parameters = { name: 'Jane' };
-  const examples = [
-    {
-      input: { content: 'Input: {{text}} from {{name}}' },
-      output: { content: 'Output: {{text}} to {{name}}' },
-    },
-  ];
-
-  const expectedResult = [
-    {
-      input: { content: 'Input: Greetings from Jane' },
-      output: { content: 'Output: Greetings to Jane' },
-    },
-  ];
-
-  t.deepEqual(palmChatPlugin.getCompiledExamples(text, parameters, examples), expectedResult);
-});
-
-test('getRequestParameters', (t) => {
-  const { palmChatPlugin } = t.context;
-  const text = 'Hello';
-  const parameters = { stream: false, name: 'John'};
-  const messages = [
-    { role: 'system', content: 'System Message' },
-    { role: 'user', content: 'Hello' },
-    { role: 'assistant', content: 'What can I do for you?' },
-    { role: 'user', content: 'Be my assistant!' },
-  ];
-  const prompt = { context: '{{text}} from {{name}}', examples: [], messages };
-
-  const requestParameters = palmChatPlugin.getRequestParameters(text, parameters, prompt);
-  const requestMessages = requestParameters.instances[0].messages;
-
-  t.is(requestMessages[0].author, 'user');
-  t.is(requestMessages[0].content, 'Hello');
-});
-
-test('getSafetyAttributes', (t) => {
-  const { palmChatPlugin } = t.context;
-  const responseData = {
-    predictions: [
-      {
-        safetyAttributes: {
-          blocked: false,
-        },
-      },
-    ],
-  };
-
-  const expectedResult = {
-    blocked: false,
-  };
-
-  t.deepEqual(palmChatPlugin.getSafetyAttributes(responseData), expectedResult);
-});
-
-test('parseResponse', (t) => {
-  const { palmChatPlugin } = t.context;
-  const responseData = {
-    predictions: [
-      {
-        candidates: [
-          {
-            content: 'Hello, how can I help you today?',
-          },
-        ],
-      },
-    ],
-  };
-
-  const expectedResult = 'Hello, how can I help you today?';
-
-  t.is(palmChatPlugin.parseResponse(responseData), expectedResult);
-});

package/tests/palmCompletionPlugin.test.js

@@ -1,58 +0,0 @@
-// palmCompletionPlugin.test.js
-
-import test from 'ava';
-import PalmCompletionPlugin from '../server/plugins/palmCompletionPlugin.js';
-import { mockPathwayResolverString } from './mocks.js';
-
-const { config, pathway, modelName, model } = mockPathwayResolverString;
-
-test.beforeEach((t) => {
-  const palmCompletionPlugin = new PalmCompletionPlugin(pathway, model);
-  t.context = { palmCompletionPlugin };
-});
-
-test('getRequestParameters', (t) => {
-  const { palmCompletionPlugin } = t.context;
-  const text = 'Hello';
-  const parameters = { stream: false, name: 'John' };
-  const prompt = {prompt:'{{text}} from {{name}}'};
-
-  const requestParameters = palmCompletionPlugin.getRequestParameters(text, parameters, prompt);
-  const requestPrompt = requestParameters.instances[0].prompt;
-
-  t.is(requestPrompt, 'Hello from John');
-});
-
-test('parseResponse', (t) => {
-  const { palmCompletionPlugin } = t.context;
-  const responseData = {
-    predictions: [
-      {
-        content: 'Hello, how can I help you today?',
-      },
-    ],
-  };
-
-  const expectedResult = 'Hello, how can I help you today?';
-
-  t.is(palmCompletionPlugin.parseResponse(responseData), expectedResult);
-});
-
-test('getSafetyAttributes', (t) => {
-  const { palmCompletionPlugin } = t.context;
-  const responseData = {
-    predictions: [
-      {
-        safetyAttributes: {
-          blocked: false,
-        },
-      },
-    ],
-  };
-
-  const expectedResult = {
-    blocked: false,
-  };
-
-  t.deepEqual(palmCompletionPlugin.getSafetyAttributes(responseData), expectedResult);
-});