@smythos/sre 1.5.1 → 1.5.4

package/src/Components/GenAILLM.class.ts

@@ -0,0 +1,395 @@
+import Joi from 'joi';
+import { IAgent as Agent } from '@sre/types/Agent.types';
+import { LLMInference } from '@sre/LLMManager/LLM.inference';
+import { TemplateString } from '@sre/helpers/TemplateString.helper';
+import { SUPPORTED_MIME_TYPES_MAP } from '@sre/constants';
+import { getMimeType } from '@sre/utils/data.utils';
+import { Component } from './Component.class';
+import { formatDataForDebug } from '@sre/utils/data.utils';
+import { AccessCandidate } from '@sre/Security/AccessControl/AccessCandidate.class';
+
+//TODO : better handling of context window exceeding max length
+
+export class GenAILLM extends Component {
+    protected schema = {
+        name: 'GenAILLM',
+        description: 'Use this component to generate a responses from an LLM',
+        settings: {
+            model: {
+                type: 'string',
+                max: 200,
+                required: true,
+            },
+            prompt: {
+                type: 'string',
+                max: 8_000_000,
+                label: 'Prompt',
+            },
+            temperature: {
+                type: 'number',
+                min: 0,
+                max: 5,
+                label: 'Temperature',
+            },
+            maxTokens: {
+                type: 'number',
+                min: 1,
+                label: 'Maximum Tokens',
+            },
+            maxThinkingTokens: {
+                type: 'number',
+                min: 1,
+                label: 'Maximum Thinking Tokens',
+            },
+            stopSequences: {
+                type: 'string',
+                max: 400,
+                label: 'Stop Sequences',
+                allowEmpty: true,
+            },
+            topP: {
+                type: 'number',
+                min: 0,
+                max: 1,
+                label: 'Top P',
+            },
+            topK: {
+                type: 'number',
+                min: 0,
+                max: 500,
+                label: 'Top K',
+            },
+            frequencyPenalty: {
+                type: 'number',
+                min: 0,
+                max: 2,
+                label: 'Frequency Penalty',
+            },
+            presencePenalty: {
+                type: 'number',
+                min: 0,
+                max: 2,
+                label: 'Presence Penalty',
+            },
+            responseFormat: {
+                type: 'string',
+                valid: ['json', 'text'],
+                label: 'Response Format',
+            },
+            passthrough: {
+                type: 'boolean',
+                description: 'If true, the LLM response will be returned as is by the agent',
+                label: 'Passthrough',
+            },
+            useSystemPrompt: {
+                type: 'boolean',
+                description: 'If true, the component will use parent agent system prompt',
+                label: 'Use System Prompt',
+            },
+            useContextWindow: {
+                type: 'boolean',
+                description: 'If true, the component will use parent agent context window',
+                label: 'Use Context Window',
+            },
+            maxContextWindowLength: {
+                type: 'number',
+                min: 0,
+                description: 'The maximum number of messages to use from this component context window (if useContextWindow is true)',
+                label: 'Maximum Context Window Length',
+            },
+        },
+        inputs: {
+            Input: {
+                type: 'Any',
+                description: 'An input that you can pass to the LLM',
+            },
+            Attachment: {
+                type: 'Binary',
+                description: 'An attachment that you can pass to the LLM',
+                optional: true,
+            },
+        },
+        outputs: {
+            Reply: {
+                default: true,
+            },
+        },
+    };
+    protected configSchema = Joi.object({
+        model: Joi.string().max(200).required(),
+        prompt: Joi.string().required().max(8_000_000).label('Prompt'), // 2M tokens is around 8M characters
+        temperature: Joi.number().min(0).max(5).label('Temperature'), // max temperature is 2 for OpenAI and togetherAI but 5 for cohere
+        maxTokens: Joi.number().min(1).label('Maximum Tokens'),
+        stopSequences: Joi.string().allow('').max(400).label('Stop Sequences'),
+        topP: Joi.number().min(0).max(1).label('Top P'),
+        topK: Joi.number().min(0).max(500).label('Top K'), // max top_k is 100 for togetherAI but 500 for cohere
+        frequencyPenalty: Joi.number().min(0).max(2).label('Frequency Penalty'),
+        presencePenalty: Joi.number().min(0).max(2).label('Presence Penalty'),
+        responseFormat: Joi.string().valid('json', 'text').optional().label('Response Format'),
+        passthrough: Joi.boolean().optional().label('Passthrough'),
+        useSystemPrompt: Joi.boolean().optional().label('Use System Prompt'),
+        useContextWindow: Joi.boolean().optional().label('Use Context Window'),
+        maxContextWindowLength: Joi.number().optional().min(0).label('Maximum Context Window Length'),
+
+        // #region Search
+        useWebSearch: Joi.boolean().optional().label('Use Search'),
+        webSearchContextSize: Joi.string().valid('high', 'medium', 'low').optional().label('Search Content Size'),
+        webSearchCity: Joi.string().max(100).optional().allow('').label('Search City'),
+        webSearchCountry: Joi.string().max(100).optional().allow('').label('Search Country'),
+        webSearchRegion: Joi.string().max(100).optional().allow('').label('Search Region'),
+        webSearchTimezone: Joi.string().max(100).optional().allow('').label('Search Timezone'),
+        // #endregion
+
+        useReasoning: Joi.boolean().optional().label('Use Reasoning'),
+        maxThinkingTokens: Joi.number().min(1).label('Maximum Thinking Tokens'),
+    });
+    constructor() {
+        super();
+    }
+    init() {}
+    async process(input, config, agent: Agent) {
+        await super.process(input, config, agent);
+
+        //let debugLog = agent.agentRuntime?.debug ? [] : undefined;
+        const logger = this.createComponentLogger(agent, config);
+
+        try {
+            logger.debug(`=== GenAILLM Log ===`);
+            let teamId = agent?.teamId;
+
+            const passThrough: boolean = config.data.passthrough || false;
+            const useContextWindow: boolean = config.data.useContextWindow || false;
+            const useSystemPrompt: boolean = config.data.useSystemPrompt || false;
+            const useWebSearch: boolean = config.data.useWebSearch || false;
+            const maxTokens: number = parseInt(config.data.maxTokens) || 1024;
+            const maxContextWindowLength: number = parseInt(config.data.maxContextWindowLength) || 1024;
+            const model: string = config.data.model || 'echo';
+            const llmInference: LLMInference = await LLMInference.getInstance(model, AccessCandidate.agent(agent.id));
+
+            // if the llm is undefined, then it means we removed the model from our system
+            if (!llmInference.connector) {
+                return {
+                    _error: `The model '${model}' is not available. Please try a different one.`,
+                    _debug: logger.output,
+                };
+            }
+
+            //const team = AccessCandidate.team(teamId);
+            //const llmRegistry = isStandardLLM ? LLMRegistry : await CustomLLMRegistry.getInstance(team);
+            const modelId = await agent.modelsProvider.getModelId(model);
+
+            logger.debug(` Model : ${modelId || model}`);
+
+            let prompt: any = TemplateString(config.data.prompt).parse(input).result;
+
+            let files: any[] = parseFiles(input, config);
+            let isMultimodalRequest = false;
+            const provider = await agent.modelsProvider.getProvider(model);
+            const isEcho = provider === 'Echo';
+
+            // Ignore files for Echo model
+            if (files?.length > 0 && !isEcho) {
+                // TODO: simplify the valid files checking logic
+                const supportedFileTypes = SUPPORTED_MIME_TYPES_MAP?.[provider] || {};
+                const modelInfo = await agent.modelsProvider.getModelInfo(model);
+                const features = modelInfo?.features || [];
+                const fileTypes = new Set(); // Set to avoid duplicates
+
+                const validFiles = await Promise.all(
+                    files.map(async (file) => {
+                        const mimeType = file?.mimetype || (await getMimeType(file));
+                        const [requestFeature = ''] =
+                            Object.entries(supportedFileTypes).find(([key, value]) => (value as string[]).includes(mimeType)) || [];
+
+                        if (mimeType) {
+                            fileTypes.add(mimeType);
+                        }
+
+                        return features?.includes(requestFeature) ? file : null;
+                    })
+                );
+
+                files = validFiles.filter(Boolean);
+
+                if (files.length === 0) {
+                    return {
+                        _error: `Model does not support ${fileTypes?.size > 0 ? Array.from(fileTypes).join(', ') : 'File(s)'}`,
+                        _debug: logger.output,
+                    };
+                }
+
+                isMultimodalRequest = true;
+            }
+
+            logger.debug(` Prompt\n`, prompt, '\n');
+
+            if (!isEcho) {
+                logger.debug(' Files\n', await Promise.all(files.map((file) => formatDataForDebug(file))));
+            }
+
+            // default to json response format
+            config.data.responseFormat = config.data?.responseFormat || 'json';
+
+            // request to LLM
+            let response: any;
+
+            const _prompt = llmInference.connector.enhancePrompt(prompt, config);
+            let messages = [];
+
+            let systemPrompt = '';
+            if (useSystemPrompt) {
+                //first we try to grab the system prompt from llmCache (in case of a ConversationHelper implementing dynamic system prompt)
+                const cachedPrompt = await agent.agentRuntime.llmCache.get('systemPrompt', 'text');
+                //if not found, we can read the system prompt from agent data.
+                systemPrompt = cachedPrompt || agent.data?.behavior || '';
+
+                if (systemPrompt) {
+                    logger.debug(' Using Agent System Prompt\n', systemPrompt);
+                }
+                if (systemPrompt) {
+                    messages = [{ role: 'system', content: systemPrompt }];
+                }
+            }
+
+            if (useContextWindow) {
+                const cachedMessages = await agent.agentRuntime.llmCache.get('messages', 'json');
+                try {
+                    const messagesJSON = typeof cachedMessages === 'string' ? JSON.parse(cachedMessages) : cachedMessages;
+                    //const contextWindow = messagesJSON.filter((message) => message.role !== 'user');
+
+                    const convMessages = await llmInference.getContextWindow(systemPrompt, messagesJSON, maxContextWindowLength, maxTokens);
+
+                    if (convMessages.length > 0) {
+                        logger.debug(` Using Agent Context Window : ${convMessages.length - 1} messages will be used`);
+                    }
+
+                    messages = [...convMessages];
+                    //messages.push(...contextWindowJSON);
+                } catch (error) {
+                    logger.warn('Error on parsing context window: ', error);
+                    console.warn(cachedMessages);
+                }
+            }
+
+            if (messages[messages.length - 1]?.role == 'user') {
+                messages[messages.length - 1].content = _prompt;
+            } else {
+                messages.push({ role: 'user', content: _prompt });
+            }
+            let finishReason = 'stop';
+            const contentPromise = new Promise(async (resolve, reject) => {
+                let _content = '';
+                let eventEmitter;
+
+                eventEmitter = await llmInference
+                    .promptStream({
+                        contextWindow: messages,
+                        files,
+                        params: {
+                            ...config.data,
+                            agentId: agent.id,
+                        },
+                    })
+                    .catch((error) => {
+                        console.error('Error on promptStream: ', error);
+                        reject(error);
+                    });
+
+                eventEmitter.on('content', (content) => {
+                    if (passThrough) {
+                        if (typeof agent.callback === 'function') {
+                            agent.callback({ content });
+                        }
+                        agent.sse.send('llm/passthrough/content', content.replace(/\n/g, '\\n'));
+                    }
+                    _content += content;
+                });
+
+                eventEmitter.on('thinking', (thinking) => {
+                    if (passThrough) {
+                        if (typeof agent.callback === 'function') {
+                            agent.callback({ thinking });
+                        }
+                        agent.sse.send('llm/passthrough/thinking', thinking.replace(/\n/g, '\\n'));
+                    }
+                });
+                eventEmitter.on('end', () => {
+                    if (passThrough) {
+                        if (typeof agent.callback === 'function') {
+                            agent.callback({ content: '\n' });
+                        }
+                        agent.sse.send('llm/passthrough/content', '\\n');
+                    }
+                    resolve(_content);
+                });
+                eventEmitter.on('interrupted', (reason) => {
+                    finishReason = reason || 'stop';
+                });
+
+                eventEmitter.on('error', (error) => {
+                    reject(error);
+                });
+            });
+            response = await contentPromise.catch((error) => {
+                return { error: error.message || error };
+            });
+            // // If the model stopped before completing the response, this is usually due to output token limit reached.
+            if (finishReason !== 'stop') {
+                return {
+                    Reply: response,
+                    _error: 'The model stopped before completing the response, this is usually due to output token limit reached.',
+                    _debug: logger.output,
+                };
+            }
+
+            // in case we have the response but it's empty string, undefined or null
+            if (!response) {
+                return { _error: ' LLM Error = Empty Response!', _debug: logger.output };
+            }
+
+            if (response?.error) {
+                const error = response?.error + ' ' + (response?.details || '');
+                logger.error(` LLM Error=`, error);
+
+                return { Output: response?.data, _error: error, _debug: logger.output };
+            }
+
+            const Reply = llmInference.connector.postProcess(response);
+            if (Reply.error) {
+                logger.error(` LLM Error=`, Reply.error);
+                return { _error: Reply.error, _debug: logger.output };
+            }
+
+            logger.debug(' Reply \n', Reply);
+
+            const result = { Reply };
+
+            result['_debug'] = logger.output;
+
+            return result;
+        } catch (error) {
+            return { _error: error.message, _debug: logger.output };
+        }
+    }
+}
+
+function parseFiles(input: any, config: any) {
+    const mediaTypes = ['Image', 'Audio', 'Video', 'Binary'];
+
+    // Parse media inputs from config
+    const inputFiles =
+        config.inputs
+            ?.filter((_input) => mediaTypes.includes(_input.type))
+            ?.flatMap((_input) => {
+                const value = input[_input.name];
+
+                if (Array.isArray(value)) {
+                    return value.map((item) => TemplateString(item).parseRaw(input).result);
+                } else {
+                    return TemplateString(value).parseRaw(input).result;
+                }
+            })
+            ?.filter((file) => file) || [];
+
+    return inputFiles;
+}