@juspay/neurolink 7.45.0 → 7.46.0

package/dist/lib/evaluation/scoring.js

@@ -0,0 +1,35 @@
+/**
+ * @file Contains the logic for mapping raw evaluation results to the structured EvaluationData type.
+ */
+/**
+ * Maps a raw `EvaluationResult` to the structured `EvaluationData` format.
+ * This includes calculating derived fields like `isOffTopic` and `alertSeverity`.
+ *
+ * @param result The raw `EvaluationResult` from the evaluator.
+ * @param threshold The score threshold to determine if the evaluation is passing.
+ * @param offTopicThreshold The score below which a response is considered off-topic.
+ * @param highSeverityThreshold The score below which a failing response is high severity.
+ * @returns A structured `EvaluationData` object.
+ */
+export function mapToEvaluationData(evalContext, result, threshold, offTopicThreshold = 5, highSeverityThreshold = 4) {
+    const isPassing = result.finalScore >= threshold;
+    return {
+        relevance: result.relevanceScore,
+        accuracy: result.accuracyScore,
+        completeness: result.completenessScore,
+        overall: result.finalScore,
+        isOffTopic: result.finalScore < offTopicThreshold,
+        alertSeverity: isPassing
+            ? "none"
+            : result.finalScore < highSeverityThreshold
+                ? "high"
+                : "medium",
+        reasoning: result.reasoning,
+        suggestedImprovements: result.suggestedImprovements,
+        evaluationModel: result.evaluationModel,
+        evaluationTime: result.evaluationTime,
+        evaluationAttempt: result.attemptNumber,
+        responseContent: evalContext.aiResponse,
+        queryContent: evalContext.userQuery,
+    };
+}

package/dist/lib/middleware/builtin/autoEvaluation.d.ts

@@ -0,0 +1,14 @@
+/**
+ * @file Implements the Auto-Evaluation Middleware for ensuring response quality.
+ */
+import type { NeuroLinkMiddleware, AutoEvaluationConfig } from "../../types/middlewareTypes.js";
+/**
+ * Creates the Auto-Evaluation middleware, which intercepts generation requests
+ * to evaluate the quality of the response. If the response quality is below a
+ * configured threshold, it can trigger retries with feedback.
+ *
+ * @param config - Configuration for the auto-evaluation middleware.
+ * @returns A `NeuroLinkMiddleware` object.
+ */
+export declare function createAutoEvaluationMiddleware(config?: AutoEvaluationConfig): NeuroLinkMiddleware;
+export default createAutoEvaluationMiddleware;

package/dist/lib/middleware/builtin/autoEvaluation.js

@@ -0,0 +1,181 @@
+/**
+ * @file Implements the Auto-Evaluation Middleware for ensuring response quality.
+ */
+import { Evaluator } from "../../evaluation/index.js";
+import { logger } from "../../utils/logger.js";
+/**
+ * Creates the Auto-Evaluation middleware, which intercepts generation requests
+ * to evaluate the quality of the response. If the response quality is below a
+ * configured threshold, it can trigger retries with feedback.
+ *
+ * @param config - Configuration for the auto-evaluation middleware.
+ * @returns A `NeuroLinkMiddleware` object.
+ */
+export function createAutoEvaluationMiddleware(config = {}) {
+    const metadata = {
+        id: "autoEvaluation",
+        name: "Auto Evaluation",
+        description: "Automatically evaluates response quality and retries if needed.",
+        priority: 90,
+        defaultEnabled: false, // Should be explicitly enabled
+    };
+    logger.debug("Auto-Evaluation Middleware Config:", config);
+    const middleware = {
+        wrapGenerate: async ({ doGenerate, params }) => {
+            const options = params;
+            const rawResult = await doGenerate();
+            const result = {
+                ...rawResult,
+                content: rawResult.text ?? "",
+                usage: {
+                    input: rawResult.usage.promptTokens,
+                    output: rawResult.usage.completionTokens,
+                    total: rawResult.usage.promptTokens + rawResult.usage.completionTokens,
+                },
+                toolCalls: rawResult.toolCalls?.map((tc) => {
+                    let parsedArgs = tc.args;
+                    if (typeof tc.args === "string") {
+                        try {
+                            parsedArgs = JSON.parse(tc.args);
+                        }
+                        catch (e) {
+                            logger.warn(`Failed to parse tool call args for tool ${tc.toolName}:`, e);
+                            parsedArgs = tc.args; // Fallback to original string if parsing fails
+                        }
+                    }
+                    return {
+                        ...tc,
+                        args: parsedArgs,
+                    };
+                }),
+            };
+            const isBlocking = config.blocking !== false;
+            if (isBlocking) {
+                const evaluationResult = await performEvaluation(config, options, result);
+                return {
+                    ...rawResult,
+                    evaluationResult,
+                };
+            }
+            else {
+                performEvaluation(config, options, result).catch((err) => {
+                    logger.error("Non-blocking auto-evaluation error:", err);
+                });
+                return rawResult;
+            }
+        },
+        wrapStream: async ({ doStream, params }) => {
+            const options = params;
+            const rawResult = await doStream();
+            const [streamForUser, streamForEvaluation] = rawResult.stream.tee();
+            // Non-blocking evaluation for streams
+            consumeAndEvaluateStream(config, options, streamForEvaluation).catch((err) => {
+                logger.error("Non-blocking stream auto-evaluation error:", err);
+            });
+            return {
+                ...rawResult,
+                stream: streamForUser,
+            };
+        },
+    };
+    return {
+        ...middleware,
+        metadata,
+    };
+}
+/**
+ * A common function to perform the evaluation logic.
+ * @param config The middleware configuration.
+ * @param options The text generation options.
+ * @param result The generation result.
+ */
+async function performEvaluation(config, options, result) {
+    const isBlocking = config.blocking !== false;
+    const threshold = config.threshold ??
+        (Number(process.env.NEUROLINK_EVALUATION_THRESHOLD) || 7);
+    try {
+        const evaluator = new Evaluator({
+            threshold,
+            provider: config.provider,
+            promptGenerator: config.promptGenerator,
+            evaluationModel: config.evaluationModel,
+        });
+        const evaluationResult = await evaluator.evaluate(options, result, threshold, config);
+        if (config.onEvaluationComplete) {
+            await config.onEvaluationComplete(evaluationResult);
+        }
+    }
+    catch (error) {
+        logger.error("Error during auto-evaluation:", error);
+        if (isBlocking) {
+            throw error;
+        }
+    }
+}
+/**
+ * Consumes a stream to build the full response and then evaluates it.
+ * @param config The middleware configuration.
+ * @param options The generation options.
+ * @param stream The stream to consume.
+ */
+async function consumeAndEvaluateStream(config, options, stream) {
+    let fullText = "";
+    let usage;
+    const toolCalls = [];
+    const reader = stream.getReader();
+    try {
+        while (true) {
+            const { done, value } = await reader.read();
+            if (done) {
+                break;
+            }
+            switch (value.type) {
+                case "text-delta":
+                    fullText += value.textDelta;
+                    break;
+                case "tool-call":
+                    {
+                        let parsedArgs;
+                        try {
+                            parsedArgs = JSON.parse(value.args);
+                        }
+                        catch (e) {
+                            logger.warn(`Failed to parse tool call args for tool ${value.toolName}:`, e);
+                            // In case of parsing failure, we can't assign a string.
+                            // Let's use an object with the raw string to maintain type safety.
+                            parsedArgs = { raw: value.args };
+                        }
+                        toolCalls.push({
+                            toolCallId: value.toolCallId,
+                            toolName: value.toolName,
+                            args: parsedArgs,
+                        });
+                    }
+                    break;
+                case "finish":
+                    usage = {
+                        input: value.usage.promptTokens,
+                        output: value.usage.completionTokens,
+                        total: value.usage.promptTokens + value.usage.completionTokens,
+                    };
+                    break;
+            }
+        }
+    }
+    finally {
+        reader.releaseLock();
+    }
+    const result = {
+        content: fullText,
+        usage,
+        toolCalls: toolCalls.length > 0 ? toolCalls : undefined,
+    };
+    // For streams, evaluation is always non-blocking from the user's perspective.
+    if (config.blocking) {
+        logger.warn("Auto-evaluation 'blocking' mode is not supported for streaming responses. Evaluation will proceed non-blockingly.");
+    }
+    // Create a new config object to force non-blocking behavior for the evaluation function
+    const nonBlockingConfig = { ...config, blocking: false };
+    await performEvaluation(nonBlockingConfig, options, result);
+}
+export default createAutoEvaluationMiddleware;

package/dist/lib/middleware/factory.js

@@ -2,6 +2,7 @@ import { wrapLanguageModel } from "ai";
 import { MiddlewareRegistry } from "./registry.js";
 import { createAnalyticsMiddleware } from "./builtin/analytics.js";
 import { createGuardrailsMiddleware } from "./builtin/guardrails.js";
+import { createAutoEvaluationMiddleware } from "./builtin/autoEvaluation.js";
 import { logger } from "../utils/logger.js";
 /**
  * Middleware factory for creating and applying middleware chains.
@@ -24,6 +25,7 @@ export class MiddlewareFactory {
         const builtInMiddlewareCreators = {
             analytics: createAnalyticsMiddleware,
             guardrails: createGuardrailsMiddleware,
+            autoEvaluation: createAutoEvaluationMiddleware,
         };
         // Register built-in presets
         this.registerPreset({
@@ -54,6 +56,7 @@ export class MiddlewareFactory {
             if (!this.registry.has(middlewareId)) {
                 const creator = builtInMiddlewareCreators[middlewareId];
                 const config = options.middlewareConfig?.[middlewareId]?.config;
+                logger.debug(`Registering built-in middleware '${middlewareId}'`, config);
                 this.registry.register(creator(config));
             }
         }
@@ -92,6 +95,7 @@ export class MiddlewareFactory {
             const middlewareConfig = this.buildMiddlewareConfig(mergedOptions);
             // Re-register middleware with the correct configuration for this call
             for (const [id, config] of Object.entries(middlewareConfig)) {
+                logger.debug(`Configuring middleware '${id}'`, { config });
                 if (config.enabled && this.registry.has(id)) {
                     const creator = this.getCreator(id);
                     if (creator) {
@@ -137,7 +141,9 @@ export class MiddlewareFactory {
         const builtInMiddlewareCreators = {
             analytics: createAnalyticsMiddleware,
             guardrails: createGuardrailsMiddleware,
+            autoEvaluation: createAutoEvaluationMiddleware,
         };
+        logger.debug("Getting creator for middleware ID:", id);
         return builtInMiddlewareCreators[id];
     }
     /**

package/dist/lib/types/evaluation.d.ts

@@ -25,6 +25,8 @@ export type EvaluationData = {
     domainAlignment?: number;
     terminologyAccuracy?: number;
     toolEffectiveness?: number;
+    responseContent?: string;
+    queryContent?: string;
     isOffTopic: boolean;
     alertSeverity: AlertSeverity;
     reasoning: string;

package/dist/lib/types/evaluationTypes.d.ts

@@ -0,0 +1,142 @@
+import type { LanguageModelV1CallOptions } from "ai";
+import type { TokenUsage } from "./analytics.js";
+import type { GenerateResult } from "./generateTypes.js";
+import type { ToolExecution } from "./tools.js";
+/**
+ * Represents the analysis of the user's query intent.
+ * This provides a basic understanding of what the user is trying to achieve.
+ */
+export interface QueryIntentAnalysis {
+    /** The type of query, e.g., asking a question or giving a command. */
+    type: "question" | "command" | "greeting" | "unknown";
+    /** The estimated complexity of the query. */
+    complexity: "low" | "medium" | "high";
+    /** Whether the query likely required the use of tools to be answered correctly. */
+    shouldHaveUsedTools: boolean;
+}
+/**
+ * Represents a single turn in an enhanced conversation history,
+ * including tool executions and evaluations for richer context.
+ */
+export interface EnhancedConversationTurn {
+    /** The role of the speaker, either 'user' or 'assistant'. */
+    role: "user" | "assistant";
+    /** The content of the message. */
+    content: string;
+    /** The timestamp of the message. */
+    timestamp: string;
+    /** Any tools that were executed as part of this turn. */
+    toolExecutions?: ToolExecution[];
+    /** The evaluation result for this turn, if applicable. */
+    evaluation?: EvaluationResult;
+}
+/**
+ * Contains all the rich context needed for a thorough, RAGAS-style evaluation.
+ * This object is constructed by the `ContextBuilder` and used by the `RAGASEvaluator`.
+ */
+export interface EnhancedEvaluationContext {
+    /** The original user query. */
+    userQuery: string;
+    /** An analysis of the user's query intent. */
+    queryAnalysis: QueryIntentAnalysis;
+    /** The AI's response that is being evaluated. */
+    aiResponse: string;
+    /** The AI provider that generated the response. */
+    provider: string;
+    /** The specific model that generated the response. */
+    model: string;
+    /** The parameters used for the generation call. */
+    generationParams: {
+        temperature?: number;
+        maxTokens?: number;
+        systemPrompt?: string;
+    };
+    /** A list of tools that were executed. */
+    toolExecutions: ToolExecution[];
+    /** The history of the conversation leading up to this turn. */
+    conversationHistory: EnhancedConversationTurn[];
+    /** The response time of the AI in milliseconds. */
+    responseTime: number;
+    /** The token usage for the generation. */
+    tokenUsage: TokenUsage;
+    /** The results of any previous evaluation attempts for this response. */
+    previousEvaluations?: EvaluationResult[];
+    /** The current attempt number for this evaluation (1-based). */
+    attemptNumber: number;
+}
+/**
+ * Represents the result of a single evaluation attempt, based on RAGAS principles.
+ */
+export interface EvaluationResult {
+    /** The final, overall score for the response, typically from 1 to 10. */
+    finalScore: number;
+    /** How well the response addresses the user's query. */
+    relevanceScore: number;
+    /** The factual accuracy of the information in the response. */
+    accuracyScore: number;
+    /** How completely the response answers the user's query. */
+    completenessScore: number;
+    /** Whether the final score meets the passing threshold. */
+    isPassing: boolean;
+    /** Constructive response from the judge LLM on how to improve the response. */
+    reasoning: string;
+    /** Specific suggestions for improving the response. */
+    suggestedImprovements: string;
+    /** The raw, unparsed response from the judge LLM. */
+    rawEvaluationResponse: string;
+    /** The model used to perform the evaluation. */
+    evaluationModel: string;
+    /** The time taken for the evaluation in milliseconds. */
+    evaluationTime: number;
+    /** The attempt number for this evaluation. */
+    attemptNumber: number;
+}
+/**
+ * Provides detailed information when a response fails quality assurance checks.
+ */
+export interface QualityErrorDetails {
+    /** The history of all evaluation attempts for this response. */
+    evaluationHistory: EvaluationResult[];
+    /** The final score of the last attempt. */
+    finalScore: number;
+    /** The total number of evaluation attempts made. */
+    attempts: number;
+    /** A summary message of the failure. */
+    message: string;
+}
+/**
+ * Configuration for the main `Evaluator` class.
+ */
+export interface EvaluationConfig {
+    /** The minimum score (1-10) for a response to be considered passing. */
+    threshold?: number;
+    /** The evaluation strategy to use. Currently only 'ragas' is supported. */
+    evaluationStrategy?: "ragas" | "custom";
+    /** The model to use for the LLM-as-judge evaluation. */
+    evaluationModel?: string;
+    /** The maximum number of evaluation attempts before failing. */
+    maxAttempts?: number;
+    /** The provider to use for the evaluation model. */
+    provider?: string;
+    /** A custom evaluator function to override the default behavior. */
+    customEvaluator?: (options: LanguageModelV1CallOptions, result: GenerateResult) => Promise<{
+        evaluationResult: EvaluationResult;
+        evalContext: EnhancedEvaluationContext;
+    }>;
+    /** The score below which a response is considered off-topic. */
+    offTopicThreshold?: number;
+    /** The score below which a failing response is considered a high severity alert. */
+    highSeverityThreshold?: number;
+    /** An optional function to generate custom evaluation prompts. */
+    promptGenerator?: GetPromptFunction;
+}
+/**
+ * A function that generates the main body of an evaluation prompt.
+ */
+export type GetPromptFunction = (context: {
+    userQuery: string;
+    history: string;
+    tools: string;
+    retryInfo: string;
+    aiResponse: string;
+}) => string;

package/dist/lib/types/evaluationTypes.js

@@ -0,0 +1 @@
+export {};

package/dist/lib/types/middlewareTypes.d.ts

@@ -1,5 +1,7 @@
 import type { LanguageModelV1Middleware } from "ai";
 import type { JsonValue } from "../types/common.js";
+import type { EvaluationData } from "./evaluation.js";
+import type { GetPromptFunction } from "./evaluationTypes.js";
 /**
  * Metadata interface for NeuroLink middleware
  * Provides additional information about middleware without affecting execution
@@ -33,7 +35,7 @@ export interface MiddlewareConfig {
     /** Whether the middleware is enabled */
     enabled?: boolean;
     /** Middleware-specific configuration */
-    config?: Record<string, JsonValue>;
+    config?: Record<string, unknown>;
     /** Conditions under which to apply this middleware */
     conditions?: MiddlewareConditions;
 }
@@ -108,7 +110,7 @@ export interface MiddlewareChainStats {
 /**
  * Built-in middleware types
  */
-export type BuiltInMiddlewareType = "analytics" | "guardrails" | "logging" | "caching" | "rateLimit" | "retry" | "timeout";
+export type BuiltInMiddlewareType = "analytics" | "guardrails" | "logging" | "caching" | "rateLimit" | "retry" | "timeout" | "autoEvaluation";
 /**
  * Middleware preset configurations
  */
@@ -144,3 +146,27 @@ export interface MiddlewareFactoryOptions {
         collectStats?: boolean;
     };
 }
+/**
+ * Configuration for the Auto-Evaluation Middleware.
+ */
+export interface AutoEvaluationConfig {
+    /** The minimum score (1-10) for a response to be considered passing. */
+    threshold?: number;
+    /** The maximum number of retry attempts before failing. */
+    maxRetries?: number;
+    /** The model to use for the LLM-as-judge evaluation. */
+    evaluationModel?: string;
+    /**
+     * If true, the middleware will wait for the evaluation to complete before returning.
+     * If the evaluation fails, it will throw an error. Defaults to true.
+     */
+    blocking?: boolean;
+    /** A callback function to be invoked with the evaluation result. */
+    onEvaluationComplete?: (evaluation: EvaluationData) => void | Promise<void>;
+    /** The score below which a response is considered off-topic. */
+    offTopicThreshold?: number;
+    /** The score below which a failing response is considered a high severity alert. */
+    highSeverityThreshold?: number;
+    promptGenerator?: GetPromptFunction;
+    provider?: string;
+}

package/dist/middleware/builtin/autoEvaluation.d.ts

@@ -0,0 +1,14 @@
+/**
+ * @file Implements the Auto-Evaluation Middleware for ensuring response quality.
+ */
+import type { NeuroLinkMiddleware, AutoEvaluationConfig } from "../../types/middlewareTypes.js";
+/**
+ * Creates the Auto-Evaluation middleware, which intercepts generation requests
+ * to evaluate the quality of the response. If the response quality is below a
+ * configured threshold, it can trigger retries with feedback.
+ *
+ * @param config - Configuration for the auto-evaluation middleware.
+ * @returns A `NeuroLinkMiddleware` object.
+ */
+export declare function createAutoEvaluationMiddleware(config?: AutoEvaluationConfig): NeuroLinkMiddleware;
+export default createAutoEvaluationMiddleware;

package/dist/middleware/builtin/autoEvaluation.js

@@ -0,0 +1,181 @@
+/**
+ * @file Implements the Auto-Evaluation Middleware for ensuring response quality.
+ */
+import { Evaluator } from "../../evaluation/index.js";
+import { logger } from "../../utils/logger.js";
+/**
+ * Creates the Auto-Evaluation middleware, which intercepts generation requests
+ * to evaluate the quality of the response. If the response quality is below a
+ * configured threshold, it can trigger retries with feedback.
+ *
+ * @param config - Configuration for the auto-evaluation middleware.
+ * @returns A `NeuroLinkMiddleware` object.
+ */
+export function createAutoEvaluationMiddleware(config = {}) {
+    const metadata = {
+        id: "autoEvaluation",
+        name: "Auto Evaluation",
+        description: "Automatically evaluates response quality and retries if needed.",
+        priority: 90,
+        defaultEnabled: false, // Should be explicitly enabled
+    };
+    logger.debug("Auto-Evaluation Middleware Config:", config);
+    const middleware = {
+        wrapGenerate: async ({ doGenerate, params }) => {
+            const options = params;
+            const rawResult = await doGenerate();
+            const result = {
+                ...rawResult,
+                content: rawResult.text ?? "",
+                usage: {
+                    input: rawResult.usage.promptTokens,
+                    output: rawResult.usage.completionTokens,
+                    total: rawResult.usage.promptTokens + rawResult.usage.completionTokens,
+                },
+                toolCalls: rawResult.toolCalls?.map((tc) => {
+                    let parsedArgs = tc.args;
+                    if (typeof tc.args === "string") {
+                        try {
+                            parsedArgs = JSON.parse(tc.args);
+                        }
+                        catch (e) {
+                            logger.warn(`Failed to parse tool call args for tool ${tc.toolName}:`, e);
+                            parsedArgs = tc.args; // Fallback to original string if parsing fails
+                        }
+                    }
+                    return {
+                        ...tc,
+                        args: parsedArgs,
+                    };
+                }),
+            };
+            const isBlocking = config.blocking !== false;
+            if (isBlocking) {
+                const evaluationResult = await performEvaluation(config, options, result);
+                return {
+                    ...rawResult,
+                    evaluationResult,
+                };
+            }
+            else {
+                performEvaluation(config, options, result).catch((err) => {
+                    logger.error("Non-blocking auto-evaluation error:", err);
+                });
+                return rawResult;
+            }
+        },
+        wrapStream: async ({ doStream, params }) => {
+            const options = params;
+            const rawResult = await doStream();
+            const [streamForUser, streamForEvaluation] = rawResult.stream.tee();
+            // Non-blocking evaluation for streams
+            consumeAndEvaluateStream(config, options, streamForEvaluation).catch((err) => {
+                logger.error("Non-blocking stream auto-evaluation error:", err);
+            });
+            return {
+                ...rawResult,
+                stream: streamForUser,
+            };
+        },
+    };
+    return {
+        ...middleware,
+        metadata,
+    };
+}
+/**
+ * A common function to perform the evaluation logic.
+ * @param config The middleware configuration.
+ * @param options The text generation options.
+ * @param result The generation result.
+ */
+async function performEvaluation(config, options, result) {
+    const isBlocking = config.blocking !== false;
+    const threshold = config.threshold ??
+        (Number(process.env.NEUROLINK_EVALUATION_THRESHOLD) || 7);
+    try {
+        const evaluator = new Evaluator({
+            threshold,
+            provider: config.provider,
+            promptGenerator: config.promptGenerator,
+            evaluationModel: config.evaluationModel,
+        });
+        const evaluationResult = await evaluator.evaluate(options, result, threshold, config);
+        if (config.onEvaluationComplete) {
+            await config.onEvaluationComplete(evaluationResult);
+        }
+    }
+    catch (error) {
+        logger.error("Error during auto-evaluation:", error);
+        if (isBlocking) {
+            throw error;
+        }
+    }
+}
+/**
+ * Consumes a stream to build the full response and then evaluates it.
+ * @param config The middleware configuration.
+ * @param options The generation options.
+ * @param stream The stream to consume.
+ */
+async function consumeAndEvaluateStream(config, options, stream) {
+    let fullText = "";
+    let usage;
+    const toolCalls = [];
+    const reader = stream.getReader();
+    try {
+        while (true) {
+            const { done, value } = await reader.read();
+            if (done) {
+                break;
+            }
+            switch (value.type) {
+                case "text-delta":
+                    fullText += value.textDelta;
+                    break;
+                case "tool-call":
+                    {
+                        let parsedArgs;
+                        try {
+                            parsedArgs = JSON.parse(value.args);
+                        }
+                        catch (e) {
+                            logger.warn(`Failed to parse tool call args for tool ${value.toolName}:`, e);
+                            // In case of parsing failure, we can't assign a string.
+                            // Let's use an object with the raw string to maintain type safety.
+                            parsedArgs = { raw: value.args };
+                        }
+                        toolCalls.push({
+                            toolCallId: value.toolCallId,
+                            toolName: value.toolName,
+                            args: parsedArgs,
+                        });
+                    }
+                    break;
+                case "finish":
+                    usage = {
+                        input: value.usage.promptTokens,
+                        output: value.usage.completionTokens,
+                        total: value.usage.promptTokens + value.usage.completionTokens,
+                    };
+                    break;
+            }
+        }
+    }
+    finally {
+        reader.releaseLock();
+    }
+    const result = {
+        content: fullText,
+        usage,
+        toolCalls: toolCalls.length > 0 ? toolCalls : undefined,
+    };
+    // For streams, evaluation is always non-blocking from the user's perspective.
+    if (config.blocking) {
+        logger.warn("Auto-evaluation 'blocking' mode is not supported for streaming responses. Evaluation will proceed non-blockingly.");
+    }
+    // Create a new config object to force non-blocking behavior for the evaluation function
+    const nonBlockingConfig = { ...config, blocking: false };
+    await performEvaluation(nonBlockingConfig, options, result);
+}
+export default createAutoEvaluationMiddleware;