@juspay/neurolink 3.0.1 → 4.1.0

package/dist/core/evaluation.js

@@ -0,0 +1,528 @@
+/**
+ * NeuroLink Unified Evaluation System
+ *
+ * Combines Universal Evaluation with Lighthouse-Enhanced capabilities
+ * - Domain-aware evaluation with sophisticated context handling
+ * - Multi-provider support with fallback strategies
+ * - Structured output with Zod schema validation
+ * - Tool usage and conversation history analysis
+ * - Enterprise-grade reliability and performance
+ */
+import { logger } from "../utils/logger.js";
+import { AIProviderFactory } from "./factory.js";
+import { z } from "zod";
+/**
+ * Unified Evaluation Schema (Lighthouse-compatible with extensions)
+ */
+export const unifiedEvaluationSchema = z.object({
+    // Core evaluation scores
+    relevanceScore: z
+        .number()
+        .min(0)
+        .max(10)
+        .describe("Score (0-10) for how well the response addresses query intent and aligns with domain/role. 10 is most relevant."),
+    accuracyScore: z
+        .number()
+        .min(0)
+        .max(10)
+        .describe("Score (0-10) for factual correctness against data, tool outputs, and domain knowledge. 10 is most accurate."),
+    completenessScore: z
+        .number()
+        .min(0)
+        .max(10)
+        .describe("Score (0-10) for how completely the response addresses the query. 10 is most complete."),
+    // Enhanced domain scores (optional)
+    domainAlignment: z
+        .number()
+        .min(0)
+        .max(10)
+        .optional()
+        .describe("Score (0-10) for how well response aligns with specified domain expertise."),
+    terminologyAccuracy: z
+        .number()
+        .min(0)
+        .max(10)
+        .optional()
+        .describe("Score (0-10) for correct usage of domain-specific terminology."),
+    toolEffectiveness: z
+        .number()
+        .min(0)
+        .max(10)
+        .optional()
+        .describe("Score (0-10) for how effectively available tools/MCPs were utilized."),
+    // Qualitative assessment
+    isOffTopic: z
+        .boolean()
+        .describe("True if the response significantly deviates from query/domain."),
+    reasoning: z
+        .string()
+        .describe("Brief justification for scores, especially if low or off-topic. Max 150 words."),
+    suggestedImprovements: z
+        .string()
+        .optional()
+        .describe("Optional: Suggestions for improving the original response. Max 100 words."),
+    alertSeverity: z
+        .enum(["low", "medium", "high", "none"])
+        .describe("Suggested alert severity considering all scores and domain context."),
+});
+/**
+ * Main unified evaluation function
+ */
+export async function performUnifiedEvaluation(context) {
+    const functionTag = "performUnifiedEvaluation";
+    const startTime = Date.now();
+    // Determine evaluation mode
+    const mode = context.mode || detectEvaluationMode(context);
+    logger.debug(`[${functionTag}] Starting unified evaluation`, {
+        mode,
+        domain: context.primaryDomain,
+        toolsUsed: context.toolsUsed?.length || 0,
+        conversationTurns: context.conversationHistory?.length || 0,
+        queryLength: context.userQuery.length,
+        responseLength: context.aiResponse.length,
+    });
+    const { parseEvaluationConfig } = await import("./evaluation-config.js");
+    const config = parseEvaluationConfig();
+    let lastError = null;
+    for (let attempt = 0; attempt <= config.retryAttempts; attempt++) {
+        try {
+            // Get evaluation model
+            const evaluationModelResult = await getEvaluationModel();
+            if (!evaluationModelResult) {
+                logger.debug(`[${functionTag}] No evaluation model available, returning defaults`);
+                return getDefaultUnifiedEvaluation("unavailable", Date.now() - startTime, context);
+            }
+            const { provider: evaluationModel, config: modelConfig } = evaluationModelResult;
+            // Create evaluation prompt based on mode
+            const evaluationPrompt = createUnifiedEvaluationPrompt(context, mode);
+            logger.debug(`[${functionTag}] Using ${mode} evaluation mode`, {
+                provider: modelConfig.providerName,
+                model: modelConfig.modelName,
+                attempt: attempt + 1,
+            });
+            // Try structured evaluation first (preferred)
+            try {
+                const structuredResult = await evaluationModel.generateObject({
+                    schema: unifiedEvaluationSchema,
+                    prompt: evaluationPrompt,
+                    temperature: 0.1,
+                    maxTokens: 1000,
+                    system: createUnifiedSystemPrompt(mode),
+                });
+                return processStructuredEvaluationResult(structuredResult.object, modelConfig, Date.now() - startTime, context, attempt + 1);
+            }
+            catch (structuredError) {
+                logger.warn(`[${functionTag}] Structured evaluation failed, using fallback`, { structuredError });
+                // Fallback to legacy generateText
+                const result = await evaluationModel.generateText({
+                    prompt: evaluationPrompt + "\n\nRespond with valid JSON only.",
+                    temperature: 0.1,
+                    maxTokens: 1000,
+                    systemPrompt: createUnifiedSystemPrompt(mode),
+                });
+                const responseText = result?.text || result?.content;
+                if (!responseText) {
+                    throw new Error("No evaluation text received from fallback");
+                }
+                return parseUnifiedEvaluationResult(responseText, modelConfig, Date.now() - startTime, context, attempt + 1);
+            }
+        }
+        catch (error) {
+            lastError = error instanceof Error ? error : new Error(String(error));
+            logger.warn(`[${functionTag}] Evaluation attempt ${attempt + 1} failed:`, lastError.message);
+            if (attempt === config.retryAttempts) {
+                break;
+            }
+            // Exponential backoff
+            await new Promise((resolve) => setTimeout(resolve, Math.pow(2, attempt) * 1000));
+        }
+    }
+    // All attempts failed
+    logger.error(`[${functionTag}] All evaluation attempts failed:`, lastError?.message);
+    return getDefaultUnifiedEvaluation(lastError?.message || "unknown-error", Date.now() - startTime, context);
+}
+/**
+ * Detect appropriate evaluation mode based on context
+ */
+function detectEvaluationMode(context) {
+    // Lighthouse mode: Has domain awareness, tool context, or conversation history
+    if (context.primaryDomain ||
+        context.toolsUsed?.length ||
+        context.conversationHistory?.length) {
+        return "lighthouse";
+    }
+    // Enhanced mode: Has rich context
+    if (context.context && Object.keys(context.context).length > 0) {
+        return "enhanced";
+    }
+    // Simple mode: Basic evaluation
+    return "simple";
+}
+/**
+ * Create unified evaluation prompt based on mode
+ */
+function createUnifiedEvaluationPrompt(context, mode) {
+    switch (mode) {
+        case "lighthouse":
+            return createLighthouseEvaluationPrompt(context);
+        case "enhanced":
+            return createEnhancedEvaluationPrompt(context);
+        case "simple":
+        default:
+            return createSimpleEvaluationPrompt(context);
+    }
+}
+/**
+ * Create Lighthouse-style domain-aware evaluation prompt
+ */
+function createLighthouseEvaluationPrompt(context) {
+    const { userQuery, aiResponse, primaryDomain = "general AI assistant", assistantRole = "AI assistant", toolContext = "No specific tools used in this interaction", conversationHistory = [], } = context;
+    const formattedHistory = formatConversationHistory(conversationHistory);
+    return `You are an AI Response Evaluator with advanced domain awareness.
+
+**EVALUATION CONTEXT**:
+
+1. **Primary Assistant Domain**: "${primaryDomain}"
+   - This defines the AI assistant's core expertise area
+   - Responses should demonstrate competency within this domain
+   - Domain-specific terminology should be used accurately
+
+2. **Assistant Role**: "${assistantRole}"
+   - This defines the specific role the assistant should fulfill
+   - Responses should align with this role's responsibilities
+
+3. **Tool Usage Context**: "${toolContext}"
+   - Tools/MCPs are capabilities the assistant used to generate the response
+   - Evaluate how effectively these tools were utilized
+   - Consider if additional tools should have been used
+
+4. **Conversation History**:
+\`\`\`
+${formattedHistory}
+\`\`\`
+
+**CRITICAL DOMAIN FAILURE ASSESSMENT**:
+Pay special attention to domain alignment. If the query is within the assistant's domain and sufficient context is available:
+- Inability to answer ("I can't help", generic errors, evasions) = HIGH ALERT
+- Incorrect domain-specific information = HIGH ALERT  
+- Misuse of domain terminology = MEDIUM-HIGH ALERT
+
+**EVALUATION CRITERIA**:
+- **relevanceScore** (0-10): Direct query addressing + domain alignment
+- **accuracyScore** (0-10): Factual correctness + terminology accuracy
+- **completenessScore** (0-10): Full query addressing + appropriate depth
+- **domainAlignment** (0-10): How well response fits the domain expertise
+- **terminologyAccuracy** (0-10): Correct use of domain-specific terms
+- **toolEffectiveness** (0-10): How well available tools were utilized
+- **isOffTopic** (boolean): True if significantly deviates from domain/query
+- **reasoning** (string): Brief explanation (max 150 words)
+- **suggestedImprovements** (string): How to improve (max 100 words)
+- **alertSeverity** ('low'|'medium'|'high'|'none'): Based on domain failure assessment
+
+**Current User Query**:
+"${userQuery}"
+
+**AI Assistant Response**:
+"${aiResponse}"
+
+Provide your assessment in the specified format.`;
+}
+/**
+ * Create enhanced evaluation prompt
+ */
+function createEnhancedEvaluationPrompt(context) {
+    const { userQuery, aiResponse, context: additionalContext } = context;
+    const contextInfo = additionalContext
+        ? `\nContext: ${JSON.stringify(additionalContext, null, 2)}`
+        : "";
+    return `Evaluate this AI response with enhanced criteria:
+
+Query: "${userQuery}"
+Response: "${aiResponse}"${contextInfo}
+
+Provide scores for:
+- relevanceScore (0-10): How well the response addresses the query
+- accuracyScore (0-10): Factual correctness and reliability
+- completenessScore (0-10): Whether the response fully answers the question
+- isOffTopic (boolean): Whether response deviates from query
+- reasoning (string): Brief explanation of scores
+- alertSeverity ('low'|'medium'|'high'|'none'): Overall quality assessment
+
+Respond in the specified format.`;
+}
+/**
+ * Create simple evaluation prompt
+ */
+function createSimpleEvaluationPrompt(context) {
+    const { userQuery, aiResponse } = context;
+    return `Rate this AI response:
+
+Q: "${userQuery}"
+A: "${aiResponse}"
+
+Provide:
+- relevanceScore (0-10)
+- accuracyScore (0-10)
+- completenessScore (0-10)
+- reasoning (brief explanation)
+
+Respond in the specified format.`;
+}
+/**
+ * Create unified system prompt based on mode
+ */
+function createUnifiedSystemPrompt(mode) {
+    const basePrompt = "You are an expert AI Response Evaluator. Respond with valid structured output only.";
+    switch (mode) {
+        case "lighthouse":
+            return `${basePrompt} Use advanced domain awareness and sophisticated context analysis for comprehensive evaluation.`;
+        case "enhanced":
+            return `${basePrompt} Consider all provided context and metadata for thorough evaluation.`;
+        case "simple":
+        default:
+            return `${basePrompt} Focus on core quality metrics: relevance, accuracy, and completeness.`;
+    }
+}
+/**
+ * Process structured evaluation result
+ */
+function processStructuredEvaluationResult(result, modelConfig, evaluationTime, context, attempt) {
+    // Calculate overall score
+    const coreScores = [
+        result.relevanceScore || 0,
+        result.accuracyScore || 0,
+        result.completenessScore || 0,
+    ];
+    const enhancedScores = [
+        result.domainAlignment,
+        result.terminologyAccuracy,
+        result.toolEffectiveness,
+    ].filter((score) => typeof score === "number" && score > 0);
+    const allScores = [...coreScores, ...enhancedScores];
+    const overall = Math.round(allScores.reduce((sum, score) => sum + score, 0) / allScores.length);
+    return {
+        // Core scores
+        relevanceScore: Math.max(0, Math.min(10, Math.round(result.relevanceScore || 0))),
+        accuracyScore: Math.max(0, Math.min(10, Math.round(result.accuracyScore || 0))),
+        completenessScore: Math.max(0, Math.min(10, Math.round(result.completenessScore || 0))),
+        overall: Math.max(0, Math.min(10, overall)),
+        // Enhanced insights
+        isOffTopic: result.isOffTopic || false,
+        alertSeverity: result.alertSeverity || "none",
+        reasoning: result.reasoning || "Evaluation completed successfully.",
+        suggestedImprovements: result.suggestedImprovements,
+        // Domain-specific scores (if available)
+        domainAlignment: result.domainAlignment
+            ? Math.max(0, Math.min(10, Math.round(result.domainAlignment)))
+            : undefined,
+        terminologyAccuracy: result.terminologyAccuracy
+            ? Math.max(0, Math.min(10, Math.round(result.terminologyAccuracy)))
+            : undefined,
+        toolEffectiveness: result.toolEffectiveness
+            ? Math.max(0, Math.min(10, Math.round(result.toolEffectiveness)))
+            : undefined,
+        // Context analysis
+        contextUtilization: {
+            conversationUsed: (context.conversationHistory?.length || 0) > 0,
+            toolsUsed: (context.toolsUsed?.length || 0) > 0,
+            domainKnowledgeUsed: !!context.primaryDomain,
+        },
+        // Enhanced metadata
+        evaluationContext: {
+            domain: context.primaryDomain || "general",
+            toolsEvaluated: context.toolsUsed || [],
+            conversationTurns: context.conversationHistory?.length || 0,
+        },
+        // Standard metadata
+        evaluationModel: `${modelConfig.providerName}/${modelConfig.modelName}`,
+        evaluationTime,
+        evaluationProvider: modelConfig.providerName,
+        evaluationAttempt: attempt,
+        evaluationConfig: {
+            mode: context.mode || "auto",
+            fallbackUsed: attempt > 1,
+            costEstimate: 0,
+        },
+    };
+}
+/**
+ * Parse evaluation result from text response
+ */
+function parseUnifiedEvaluationResult(evaluationText, modelConfig, evaluationTime, context, attempt) {
+    try {
+        // Clean and parse JSON
+        const cleanText = evaluationText.trim().replace(/```json\s*|```\s*/g, "");
+        const jsonMatch = cleanText.match(/\{[^]*?\}/s);
+        if (jsonMatch) {
+            const parsed = JSON.parse(jsonMatch[0]);
+            return processStructuredEvaluationResult(parsed, modelConfig, evaluationTime, context, attempt);
+        }
+        // Fallback to regex parsing with improved patterns
+        const relevanceMatch = evaluationText.match(/(?:relevance[Score"\s]*:?["\s]*(\d+)|Relevance["\s]*:?["\s]*(\d+)|relevance.*?(\d+))/i);
+        const accuracyMatch = evaluationText.match(/(?:accuracy[Score"\s]*:?["\s]*(\d+)|Accuracy["\s]*:?["\s]*(\d+)|accuracy.*?(\d+))/i);
+        const completenessMatch = evaluationText.match(/(?:completeness[Score"\s]*:?["\s]*(\d+)|Completeness["\s]*:?["\s]*(\d+)|completeness.*?(\d+))/i);
+        // Extract scores with fallback to default values
+        const relevance = relevanceMatch
+            ? parseInt(relevanceMatch[1] || relevanceMatch[2] || relevanceMatch[3], 10)
+            : 8; // Default fallback score
+        const accuracy = accuracyMatch
+            ? parseInt(accuracyMatch[1] || accuracyMatch[2] || accuracyMatch[3], 10)
+            : 8; // Default fallback score
+        const completeness = completenessMatch
+            ? parseInt(completenessMatch[1] || completenessMatch[2] || completenessMatch[3], 10)
+            : 8; // Default fallback score
+        return {
+            relevanceScore: Math.max(0, Math.min(10, relevance)),
+            accuracyScore: Math.max(0, Math.min(10, accuracy)),
+            completenessScore: Math.max(0, Math.min(10, completeness)),
+            overall: Math.round((relevance + accuracy + completeness) / 3),
+            isOffTopic: false,
+            alertSeverity: "none",
+            reasoning: "Parsed using regex fallback - response was not in expected JSON format.",
+            evaluationModel: `${modelConfig.providerName}/${modelConfig.modelName}`,
+            evaluationTime,
+            evaluationProvider: modelConfig.providerName,
+            evaluationAttempt: attempt,
+            evaluationConfig: {
+                mode: "fallback",
+                fallbackUsed: true,
+                costEstimate: 0,
+            },
+        };
+    }
+    catch (error) {
+        logger.error("Failed to parse unified evaluation result", { error });
+        return getDefaultUnifiedEvaluation("parse-error", evaluationTime, context);
+    }
+}
+/**
+ * Get default evaluation when evaluation fails
+ */
+function getDefaultUnifiedEvaluation(reason, evaluationTime, context) {
+    return {
+        relevanceScore: 0,
+        accuracyScore: 0,
+        completenessScore: 0,
+        overall: 0,
+        isOffTopic: false,
+        alertSeverity: "high",
+        reasoning: `Evaluation unavailable (${reason}). This may be due to missing API keys, network issues, or service unavailability.`,
+        suggestedImprovements: "Check evaluation system configuration, API credentials, and network connectivity.",
+        evaluationModel: "unavailable",
+        evaluationTime,
+        evaluationProvider: "none",
+        evaluationAttempt: 0,
+        evaluationConfig: {
+            mode: "default",
+            fallbackUsed: true,
+            costEstimate: 0,
+        },
+        contextUtilization: {
+            conversationUsed: (context.conversationHistory?.length || 0) > 0,
+            toolsUsed: (context.toolsUsed?.length || 0) > 0,
+            domainKnowledgeUsed: !!context.primaryDomain,
+        },
+        evaluationContext: {
+            domain: context.primaryDomain || "unknown",
+            toolsEvaluated: context.toolsUsed || [],
+            conversationTurns: context.conversationHistory?.length || 0,
+        },
+    };
+}
+/**
+ * Enhanced evaluation model selection
+ */
+export async function getEvaluationModel() {
+    const { parseEvaluationConfig, getProviderFallbackOrder } = await import("./evaluation-config.js");
+    const { getProviderConfig } = await import("./evaluation-providers.js");
+    const config = parseEvaluationConfig();
+    const fallbackOrder = getProviderFallbackOrder(config);
+    for (const providerName of fallbackOrder) {
+        try {
+            const providerConfig = getProviderConfig(providerName);
+            if (!providerConfig) {
+                continue;
+            }
+            let modelName = config.model;
+            if (modelName === "auto" || !config.model) {
+                modelName =
+                    providerConfig.models[config.mode] || providerConfig.models.fast;
+            }
+            const provider = await AIProviderFactory.createProvider(providerName, modelName);
+            if (provider) {
+                return {
+                    provider,
+                    config: {
+                        providerName,
+                        modelName,
+                        providerConfig,
+                        evaluationConfig: config,
+                    },
+                };
+            }
+        }
+        catch (error) {
+            if (!config.fallbackEnabled) {
+                throw error;
+            }
+            continue;
+        }
+    }
+    return null;
+}
+/**
+ * Format conversation history for evaluation
+ */
+function formatConversationHistory(history) {
+    if (!history?.length) {
+        return "No prior conversation context.";
+    }
+    return history
+        .slice(-3) // Last 3 turns
+        .map((msg, i) => `${i + 1}. ${msg.role.toUpperCase()}: ${msg.content.substring(0, 200)}${msg.content.length > 200 ? "..." : ""}`)
+        .join("\n");
+}
+/**
+ * Create simple evaluation context (backward compatibility)
+ */
+export function createSimpleEvaluationContext(prompt, response, context) {
+    return {
+        userQuery: prompt,
+        aiResponse: response,
+        context,
+        mode: "simple",
+    };
+}
+/**
+ * Create enhanced evaluation context
+ */
+export function createEnhancedEvaluationContext(userQuery, aiResponse, options = {}) {
+    return {
+        userQuery,
+        aiResponse,
+        primaryDomain: options.domain,
+        assistantRole: options.role,
+        toolsUsed: options.toolsUsed,
+        toolContext: options.toolsUsed?.length
+            ? `Tools used: ${options.toolsUsed.join(", ")}`
+            : undefined,
+        conversationHistory: options.conversationHistory,
+        sessionId: options.sessionId,
+        context: options.context,
+        mode: "lighthouse",
+    };
+}
+// Legacy compatibility wrapper for old function signature
+export async function evaluateResponse(prompt, response, context, evaluationDomain, toolUsageContext, conversationHistory) {
+    // Convert old arguments to new context format
+    const unifiedContext = {
+        userQuery: prompt,
+        aiResponse: response,
+        context,
+        primaryDomain: evaluationDomain,
+        toolContext: toolUsageContext,
+        conversationHistory: conversationHistory,
+        mode: evaluationDomain ? "lighthouse" : "simple",
+    };
+    return performUnifiedEvaluation(unifiedContext);
+}

package/dist/core/factory.js

@@ -81,31 +81,39 @@ export class AIProviderFactory {
             //     error: dynamicError instanceof Error ? dynamicError.message : String(dynamicError),
             //   });
             // }
-            // Resolve dynamic model if available
-            let resolvedModelName = modelName;
-            if (!modelName || modelName === "default") {
-                try {
-                    const normalizedProvider = this.normalizeProviderName(providerName);
-                    const dynamicModel = dynamicModelProvider.resolveModel(normalizedProvider, modelName || undefined);
-                    if (dynamicModel) {
-                        resolvedModelName = dynamicModel.id;
-                        logger.debug(`[${functionTag}] Resolved dynamic model`, {
-                            provider: normalizedProvider,
-                            requestedModel: modelName || "default",
-                            resolvedModel: resolvedModelName,
-                            displayName: dynamicModel.displayName,
-                            pricing: dynamicModel.pricing.input,
-                        });
-                    }
-                }
-                catch (resolveError) {
-                    logger.debug(`[${functionTag}] Dynamic model resolution failed, using fallback`, {
-                        error: resolveError instanceof Error
-                            ? resolveError.message
-                            : String(resolveError),
-                    });
-                }
-            }
+            // COMPREHENSIVE FIX: Disable dynamic model resolution completely until provider is fixed
+            // This prevents stale gemini-1.5-pro-latest from overriding correct gemini-2.5-pro defaults
+            const resolvedModelName = modelName;
+            // COMMENTED OUT: Dynamic model resolution causing 1.5 vs 2.5 Pro issues
+            // if (!modelName || modelName === "default") {
+            //   try {
+            //     const normalizedProvider = this.normalizeProviderName(providerName);
+            //     const dynamicModel = dynamicModelProvider.resolveModel(
+            //       normalizedProvider,
+            //       modelName || undefined,
+            //     );
+            //     if (dynamicModel) {
+            //       resolvedModelName = dynamicModel.id;
+            //       logger.debug(`[${functionTag}] Resolved dynamic model`, {
+            //         provider: normalizedProvider,
+            //         requestedModel: modelName || "default",
+            //         resolvedModel: resolvedModelName,
+            //         displayName: dynamicModel.displayName,
+            //         pricing: dynamicModel.pricing.input,
+            //       });
+            //     }
+            //   } catch (resolveError) {
+            //     logger.debug(
+            //       `[${functionTag}] Dynamic model resolution failed, using fallback`,
+            //       {
+            //         error:
+            //           resolveError instanceof Error
+            //             ? resolveError.message
+            //             : String(resolveError),
+            //       },
+            //     );
+            //   }
+            // }
             let provider;
             switch (providerName.toLowerCase()) {
                 case "vertex":