@juspay/neurolink 9.75.0 → 9.77.0

package/dist/lib/core/toolDedup.js

@@ -0,0 +1,194 @@
+/**
+ * Tool-signature deduplication — pure module, no I/O, no side effects.
+ *
+ * ## Algorithm
+ *
+ * `computeToolSignature` builds a canonical, order-insensitive string from:
+ *   1. The tool's name (exact, case-preserved).
+ *   2. A normalised description: lowercased, all whitespace runs collapsed to a
+ *      single space, leading/trailing whitespace stripped.
+ *   3. A sorted, deduplicated list of JSON-schema property names (top-level
+ *      `properties` keys from `inputSchema`).  If the schema exposes a `type`
+ *      field on each property, that type string is appended as `name:type`.
+ *
+ * `dedupeTools` clusters tools by pairwise token-set Jaccard similarity:
+ *
+ *   Jaccard(A, B) = |A ∩ B| / |A ∪ B|
+ *
+ * where A and B are the _sets_ of whitespace-split tokens from each tool's
+ * canonical signature.  This is deterministic, symmetric, and independent of
+ * token ordering — making it robust to minor rewording while being fast (O(n²)
+ * in the number of tools, which is bounded in practice by provider limits).
+ *
+ * The first tool encountered in stable input-iteration order becomes the
+ * cluster representative; subsequent tools in the same cluster are dropped.
+ */
+// ---------------------------------------------------------------------------
+// Internal helpers
+// ---------------------------------------------------------------------------
+/**
+ * Extract top-level property names (and optionally their types) from a JSON
+ * Schema-shaped `inputSchema`.  Returns a sorted, stable list.
+ *
+ * The `inputSchema` on a `Tool` is typed as `FlexibleSchema<INPUT>` (opaque at
+ * runtime), so we access it via index to avoid casting.
+ */
+function extractSchemaTokens(inputSchema) {
+    if (!inputSchema) {
+        return [];
+    }
+    // FlexibleSchema may be a Zod schema, a jsonSchema() wrapper, or a plain
+    // JSON Schema object.  All three expose `jsonSchema` or fall through to the
+    // raw object.  We look for the plain JSON Schema `properties` key.
+    const schema = inputSchema;
+    // Vercel AI SDK wraps plain schemas in { jsonSchema: {...}, ... }
+    const rawSchema = schema["jsonSchema"] !== undefined
+        ? schema["jsonSchema"]
+        : schema;
+    const properties = rawSchema["properties"];
+    if (!properties || typeof properties !== "object" || properties === null) {
+        return [];
+    }
+    const tokens = [];
+    const props = properties;
+    for (const propName of Object.keys(props)) {
+        const propDef = props[propName];
+        if (propDef && typeof propDef === "object" && propDef !== null) {
+            const t = propDef["type"];
+            tokens.push(typeof t === "string" ? `${propName}:${t}` : propName);
+        }
+        else {
+            tokens.push(propName);
+        }
+    }
+    return tokens.sort();
+}
+/**
+ * Normalise a description string: lowercase, collapse whitespace, strip ends.
+ */
+function normaliseDescription(desc) {
+    if (!desc) {
+        return "";
+    }
+    return desc.toLowerCase().replace(/\s+/g, " ").trim();
+}
+// ---------------------------------------------------------------------------
+// Public API
+// ---------------------------------------------------------------------------
+/**
+ * Build a canonical, order-insensitive signature string for a named tool.
+ *
+ * The signature is composed of:
+ *   - the tool's name
+ *   - a normalised description (lowercased, whitespace-collapsed)
+ *   - sorted parameter property names (with types when available)
+ *
+ * Stable regardless of property declaration order in the schema.
+ */
+export function computeToolSignature(name, tool) {
+    const descPart = normaliseDescription(tool.description);
+    const schemaParts = extractSchemaTokens(tool.inputSchema);
+    // Pipe-separated sections keep tokens within each section unambiguous.
+    return [name, descPart, schemaParts.join(",")].join("|");
+}
+/**
+ * Compute the token-set Jaccard similarity between two signature strings.
+ *
+ *   Jaccard(A, B) = |A ∩ B| / |A ∪ B|
+ *
+ * Splits on whitespace and also on the pipe/comma delimiters used by
+ * `computeToolSignature` so that structural differences (e.g. a parameter
+ * missing from one tool) are properly reflected.
+ */
+function jaccardSimilarity(sigA, sigB) {
+    const tokenise = (s) => new Set(s.split(/[\s|,]+/).filter((t) => t.length > 0));
+    const setA = tokenise(sigA);
+    const setB = tokenise(sigB);
+    if (setA.size === 0 && setB.size === 0) {
+        return 1;
+    }
+    if (setA.size === 0 || setB.size === 0) {
+        return 0;
+    }
+    let intersection = 0;
+    for (const token of setA) {
+        if (setB.has(token)) {
+            intersection += 1;
+        }
+    }
+    const union = setA.size + setB.size - intersection;
+    return intersection / union;
+}
+/**
+ * Collapse near-duplicate tools in a name→Tool record.
+ *
+ * When `options.enabled` is falsy (the default), returns the original record
+ * and an empty `removed` array — byte-for-byte unchanged behaviour.
+ *
+ * When enabled, tools whose token-set Jaccard similarity (over their canonical
+ * signatures) meets or exceeds `options.threshold` (default 0.9) are collapsed
+ * to a single representative (the first in stable iteration order).
+ *
+ * Any exception thrown internally returns the ORIGINAL tool set (fail-open).
+ */
+export function dedupeTools(tools, options) {
+    const noOp = { tools, removed: [] };
+    if (!options.enabled) {
+        return noOp;
+    }
+    try {
+        const rawThreshold = options.threshold ?? 0.9;
+        const threshold = Math.min(1, Math.max(0, rawThreshold));
+        const entries = Object.entries(tools);
+        if (entries.length <= 1) {
+            return noOp;
+        }
+        // Pre-compute signatures once.
+        const signatures = new Map();
+        for (const [name, tool] of entries) {
+            signatures.set(name, computeToolSignature(name, tool));
+        }
+        // representative[toolName] = name of the cluster representative.
+        // Tools not yet assigned to a cluster act as their own representative.
+        const representativeOf = new Map();
+        for (const [nameA] of entries) {
+            if (representativeOf.has(nameA)) {
+                // Already absorbed into another cluster.
+                continue;
+            }
+            const sigA = signatures.get(nameA) ?? "";
+            for (const [nameB] of entries) {
+                if (nameA === nameB || representativeOf.has(nameB)) {
+                    continue;
+                }
+                const sigB = signatures.get(nameB) ?? "";
+                const sim = jaccardSimilarity(sigA, sigB);
+                if (sim >= threshold) {
+                    representativeOf.set(nameB, nameA);
+                }
+            }
+        }
+        if (representativeOf.size === 0) {
+            return noOp;
+        }
+        const dedupedTools = {};
+        const removed = [];
+        for (const [name, tool] of entries) {
+            const rep = representativeOf.get(name);
+            if (rep !== undefined) {
+                // Compute similarity again to include in the removed record.
+                const sim = jaccardSimilarity(signatures.get(name) ?? "", signatures.get(rep) ?? "");
+                removed.push({ name, duplicateOf: rep, similarity: sim });
+            }
+            else {
+                dedupedTools[name] = tool;
+            }
+        }
+        return { tools: dedupedTools, removed };
+    }
+    catch {
+        // Fail open — return the original set.
+        return noOp;
+    }
+}
+//# sourceMappingURL=toolDedup.js.map

package/dist/lib/index.d.ts

@@ -214,6 +214,7 @@ export declare function createBestAIProvider(requestedProvider?: string, modelNa
  * ```
  */
 export { CircuitBreakerManager, calculateExpiresAt, createMCPServer, createOAuthProviderFromConfig, DEFAULT_HTTP_RETRY_CONFIG, DEFAULT_RATE_LIMIT_CONFIG, executeMCP, FileTokenStorage, getMCPStats, getServerInfo, globalCircuitBreakerManager, globalRateLimiterManager, HTTPRateLimiter, InMemoryTokenStorage, initializeMCPEcosystem, isRetryableHTTPError, isRetryableStatusCode, isTokenExpired, listMCPs, CircuitBreakerOpenError, MCPCircuitBreaker, mcpLogger, NeuroLinkOAuthProvider, RateLimiterManager, validateServerTools, validateTool as validateMCPTool, withHTTPRetry, MCPToolRegistry, ExternalServerManager, MCPClientFactory, ToolRouter, createToolRouter, DEFAULT_ROUTER_CONFIG, ToolCache, createToolCache, DEFAULT_CACHE_CONFIG, ToolResultCache, createToolResultCache, RequestBatcher, ToolCallBatcher, createRequestBatcher, createToolCallBatcher, DEFAULT_BATCH_CONFIG, inferAnnotations, createAnnotatedTool, validateAnnotations, filterToolsByAnnotations, mergeAnnotations, getAnnotationSummary, requiresConfirmation, isSafeToRetry, getToolSafetyLevel, ElicitationManager, globalElicitationManager, EnhancedToolDiscovery, MCPRegistryClient, globalMCPRegistryClient, getWellKnownServer, getAllWellKnownServers, MCPServerBase, MultiServerManager, globalMultiServerManager, AgentExposureManager, exposeAgentAsTool, exposeAgentsAsTools, exposeWorkflowAsTool, exposeWorkflowsAsTools, globalAgentExposureManager, ServerCapabilitiesManager, createTextResource, createJsonResource, createPrompt, neuroLinkToolToMCP, mcpToolToNeuroLink, batchConvertToMCP, batchConvertToNeuroLink, sanitizeToolName, validateToolName, createToolFromFunction, mcpProtocolToolToServerTool, serverToolToMCPProtocol, TOOL_COMPATIBILITY, ToolIntegrationManager, globalToolIntegrationManager, wrapToolWithElicitation, wrapToolsWithElicitation, createElicitationContext, confirmationMiddleware, validationMiddleware, loggingMiddleware, createRetryMiddleware, createTimeoutMiddleware, createToolMiddlewareChain, ElicitationProtocolAdapter, globalElicitationProtocol, isElicitationProtocolMessage, protocolMessageToElicitation, elicitationResponseToProtocol, createElicitationRequest, createElicitationResponse, createElicitationCancel, createTextInputRequest, createSelectRequest, createConfirmationRequest, createFormRequest, } from "./mcp/index.js";
+export { computeToolSignature, dedupeTools } from "./core/toolDedup.js";
 export { logger } from "./utils/logger.js";
 export { getPoolStats } from "./utils/redis.js";
 export { ToolRoutingCache } from "./core/toolRoutingCache.js";
@@ -336,6 +337,36 @@ export { BALANCED_ADAPTIVE_WORKFLOW, createAdaptiveWorkflow, QUALITY_MAX_WORKFLO
 export { CONSENSUS_3_FAST_WORKFLOW, CONSENSUS_3_WORKFLOW, createConsensus3WithPrompt, } from "./workflow/workflows/consensusWorkflow.js";
 export { AGGRESSIVE_FALLBACK_WORKFLOW, FAST_FALLBACK_WORKFLOW, } from "./workflow/workflows/fallbackWorkflow.js";
 export { createMultiJudgeWorkflow, MULTI_JUDGE_3_WORKFLOW, MULTI_JUDGE_5_WORKFLOW, } from "./workflow/workflows/multiJudgeWorkflow.js";
+/**
+ * ModelPool and RequestRouter — opt-in multi-provider failover with
+ * error-class-aware cooldown, and a pluggable pre-call provider/model router.
+ *
+ * @example ModelPool
+ * ```typescript
+ * import { ModelPool, classifyProviderError } from '@juspay/neurolink';
+ *
+ * const pool = new ModelPool({
+ *   members: [
+ *     { provider: 'anthropic', model: 'claude-sonnet-4-5' },
+ *     { provider: 'vertex',    model: 'gemini-2.5-flash' },
+ *   ],
+ *   strategy: 'priority',
+ *   cooldownMs: 30_000,
+ * });
+ * ```
+ *
+ * @example RequestRouter
+ * ```typescript
+ * import { createDefaultRequestRouter } from '@juspay/neurolink';
+ *
+ * const router = createDefaultRequestRouter({
+ *   visionTier: { provider: 'vertex', model: 'gemini-2.5-pro' },
+ *   largeTier:  { provider: 'anthropic', model: 'claude-opus-4-5' },
+ *   smallTier:  { provider: 'anthropic', model: 'claude-haiku-3-5' },
+ * });
+ * ```
+ */
+export { classifyProviderError, ModelPool, createDefaultRequestRouter, } from "./routing/index.js";
 /**
  * Server Adapters for exposing NeuroLink as HTTP APIs
  *

package/dist/lib/index.js

@@ -354,6 +354,8 @@ neuroLinkToolToMCP, mcpToolToNeuroLink, batchConvertToMCP, batchConvertToNeuroLi
 ToolIntegrationManager, globalToolIntegrationManager, wrapToolWithElicitation, wrapToolsWithElicitation, createElicitationContext, confirmationMiddleware, validationMiddleware, loggingMiddleware, createRetryMiddleware, createTimeoutMiddleware, createToolMiddlewareChain, 
 // MCP Enhancements - Elicitation Protocol
 ElicitationProtocolAdapter, globalElicitationProtocol, isElicitationProtocolMessage, protocolMessageToElicitation, elicitationResponseToProtocol, createElicitationRequest, createElicitationResponse, createElicitationCancel, createTextInputRequest, createSelectRequest, createConfirmationRequest, createFormRequest, } from "./mcp/index.js";
+// Tool signature deduplication utilities (opt-in, fail-open)
+export { computeToolSignature, dedupeTools } from "./core/toolDedup.js";
 export { logger } from "./utils/logger.js";
 export { getPoolStats } from "./utils/redis.js";
 // Pre-call tool routing cache (ITEM C: LRU+TTL cache + session stickiness)
@@ -539,6 +541,39 @@ export { AGGRESSIVE_FALLBACK_WORKFLOW, FAST_FALLBACK_WORKFLOW, } from "./workflo
 // Pre-built workflows - Multi-judge
 export { createMultiJudgeWorkflow, MULTI_JUDGE_3_WORKFLOW, MULTI_JUDGE_5_WORKFLOW, } from "./workflow/workflows/multiJudgeWorkflow.js";
 // ============================================================================
+// MODEL POOL & REQUEST ROUTER - Pluggable pre-call routing + failover
+// ============================================================================
+/**
+ * ModelPool and RequestRouter — opt-in multi-provider failover with
+ * error-class-aware cooldown, and a pluggable pre-call provider/model router.
+ *
+ * @example ModelPool
+ * ```typescript
+ * import { ModelPool, classifyProviderError } from '@juspay/neurolink';
+ *
+ * const pool = new ModelPool({
+ *   members: [
+ *     { provider: 'anthropic', model: 'claude-sonnet-4-5' },
+ *     { provider: 'vertex',    model: 'gemini-2.5-flash' },
+ *   ],
+ *   strategy: 'priority',
+ *   cooldownMs: 30_000,
+ * });
+ * ```
+ *
+ * @example RequestRouter
+ * ```typescript
+ * import { createDefaultRequestRouter } from '@juspay/neurolink';
+ *
+ * const router = createDefaultRequestRouter({
+ *   visionTier: { provider: 'vertex', model: 'gemini-2.5-pro' },
+ *   largeTier:  { provider: 'anthropic', model: 'claude-opus-4-5' },
+ *   smallTier:  { provider: 'anthropic', model: 'claude-haiku-3-5' },
+ * });
+ * ```
+ */
+export { classifyProviderError, ModelPool, createDefaultRequestRouter, } from "./routing/index.js";
+// ============================================================================
 // SERVER ADAPTERS - HTTP API Framework Integration
 // ============================================================================
 // Server Types

package/dist/lib/neurolink.d.ts

@@ -5,7 +5,7 @@
  * Enhanced AI provider system with natural MCP tool access.
  * Uses real MCP infrastructure for tool discovery and execution.
  */
-import type { CompactionConfig, CompactionResult, SpanData, ObservabilityConfig, MetricsSummary, MCPToolAnnotations, TraceView, AuthenticatedContext, AuthProvider, JsonObject, NeuroLinkEvents, TypedEventEmitter, MCPEnhancementsConfig, NeuroLinkAuthConfig, NeurolinkConstructorConfig, ChatMessage, ExternalMCPOperationResult, ExternalMCPServerInstance, ExternalMCPToolInfo, GenerateOptions, GenerateResult, ProviderStatus, TextGenerationOptions, TextGenerationResult, MCPExecutableTool, MCPServerInfo, MCPStatus, StreamOptions, StreamResult, ToolExecutionContext, ToolExecutionSummary, ToolInfo, ToolRegistrationOptions, BatchOperationResult, StreamGenerationEndContext, ToolRoutingServerDescriptor } from "./types/index.js";
+import type { CompactionConfig, CompactionResult, SpanData, ObservabilityConfig, MetricsSummary, MCPToolAnnotations, TraceView, AuthenticatedContext, AuthProvider, JsonObject, NeuroLinkEvents, TypedEventEmitter, MCPEnhancementsConfig, NeuroLinkAuthConfig, NeurolinkConstructorConfig, ChatMessage, ExternalMCPOperationResult, ExternalMCPServerInstance, ExternalMCPToolInfo, GenerateOptions, GenerateResult, ProviderStatus, TextGenerationOptions, TextGenerationResult, MCPExecutableTool, MCPServerInfo, MCPStatus, StreamOptions, StreamResult, ToolExecutionContext, ToolExecutionSummary, ToolInfo, ToolRegistrationOptions, BatchOperationResult, StreamGenerationEndContext, ToolRoutingServerDescriptor, ToolDedupConfig } from "./types/index.js";
 import { ConversationMemoryManager } from "./core/conversationMemoryManager.js";
 import type { RedisConversationMemoryManager } from "./core/redisConversationMemoryManager.js";
 import { ExternalServerManager } from "./mcp/externalServerManager.js";
@@ -102,12 +102,15 @@ export declare class NeuroLink {
     private conversationMemoryConfig?;
     private toolRoutingConfig?;
     private toolRoutingCacheInstance?;
+    private toolDedupConfig?;
     private enableOrchestration;
     private authProvider?;
     private pendingAuthConfig?;
     private authInitPromise?;
     private credentials?;
     private readonly fallbackConfig;
+    private readonly modelPool;
+    private readonly requestRouter;
     /**
      * Merge instance-level credentials with per-call credentials.
      *
@@ -611,6 +614,25 @@ export declare class NeuroLink {
     private maybeHandleEarlyGenerateResult;
     private runStandardGenerateRequest;
     private maybeApplyGenerateOrchestration;
+    /**
+     * Applies the host-configured `requestRouter` to `options` in place.
+     *
+     * The router is skipped when:
+     *   - no `requestRouter` is configured on this instance, or
+     *   - the caller explicitly set both `options.provider` AND `options.model`
+     *     (we only skip the router if BOTH are set; a caller setting only one
+     *     still lets the router fill in the other field).
+     *
+     * Fails open: any router error is logged at WARN level and the call
+     * continues with the original options unmodified.
+     *
+     * @param options — the mutable options object (generate or stream).
+     * @param promptText — the text prompt used to build the RouterInputContext.
+     * @param hasTools — true when at least one tool is available for this call.
+     * @param requiresVision — true when the call includes image attachments.
+     * @param thinkingLevel — optional thinking level string from the call options.
+     */
+    private applyRequestRouter;
     private prepareGenerateAugmentations;
     private buildGenerateTextOptions;
     private finalizeGenerateRequestResult;
@@ -1089,6 +1111,19 @@ export declare class NeuroLink {
      * @see {@link NeuroLink.executeTool} for events related to tool execution
      */
     getEventEmitter(): TypedEventEmitter<NeuroLinkEvents>;
+    /**
+     * Returns the instance-level tool-dedup configuration, or `undefined` when
+     * toolDedup was not provided at construction time.
+     *
+     * The stored object is returned as-is whenever `toolDedup` was supplied,
+     * including when `enabled: false` — only the complete absence of a `toolDedup`
+     * option results in `undefined`.
+     *
+     * Called by `BaseProvider.applyToolFiltering` so the dedup pass uses the
+     * same config for every generate/stream call without threading an extra
+     * parameter through the full call stack.
+     */
+    getToolDedupConfig(): ToolDedupConfig | undefined;
     /**
      * Curator P1-1: synchronous credential health check for a single provider.
      *