@juspay/neurolink 12.14.21 → 12.15.1

package/dist/types/proxy.d.ts

@@ -221,6 +221,7 @@ export type InternalResult = {
         total: number;
         cacheCreationTokens?: number;
         cacheReadTokens?: number;
+        reasoning?: number;
     };
     toolCalls?: Array<{
         toolCallId: string;
@@ -576,6 +577,8 @@ export type RequestLogEntry = {
     method: string;
     path: string;
     model: string;
+    /** Client model when a different upstream model served the request. */
+    requestedModel?: string;
     stream: boolean;
     toolCount: number;
     account: string;
@@ -590,10 +593,14 @@ export type RequestLogEntry = {
     errorCode?: string;
     /** Whether changing credentials can affect this transport failure. */
     transportScope?: ProxyTransportScope;
+    /** True when input includes the cache breakdown (native Codex wire usage). */
+    inputIncludesCachedTokens?: boolean;
     inputTokens?: number;
     outputTokens?: number;
     cacheCreationTokens?: number;
     cacheReadTokens?: number;
+    /** Reasoning tokens are a subset of output, never additional usage. */
+    reasoningTokens?: number;
     /**
      * Provider that actually served the request, for costing. Absent on records
      * written before this field existed; `proxyAnalysis` then falls back to a
@@ -814,6 +821,10 @@ export type RequestAttemptLogEntry = {
     rateLimitKind?: "transient" | "quota";
     /** Reset-aware cooldown reason selected for a rate-limited attempt. */
     cooldownReason?: "transient" | "session" | "weekly" | "unified";
+    /** Reset reported by a structured quota rejection, epoch milliseconds. */
+    quotaResetAt?: number;
+    /** Unknown scope never implies an account-wide long-window rejection. */
+    quotaScope?: "session" | "weekly" | "unknown";
     inputTokens?: number;
     outputTokens?: number;
     cacheCreationTokens?: number;
@@ -828,9 +839,9 @@ export type RequestAttemptLogEntry = {
     traceFlags?: number;
 };
 /** Additional fields recorded when a Codex response becomes client-final. */
-export type CodexFinalLogExtra = Partial<Pick<RequestLogEntry, "errorType" | "errorMessage" | "errorCode" | "transportScope" | "inputTokens" | "outputTokens" | "cacheReadTokens" | "cacheCreationTokens" | "terminalOutcome" | "firstUsefulOutputMs" | "firstUsefulOutputStatus" | "firstUsefulOutputEvent">>;
+export type CodexFinalLogExtra = Partial<Pick<RequestLogEntry, "errorType" | "errorMessage" | "errorCode" | "transportScope" | "inputTokens" | "outputTokens" | "cacheReadTokens" | "cacheCreationTokens" | "reasoningTokens" | "terminalOutcome" | "firstUsefulOutputMs" | "firstUsefulOutputStatus" | "firstUsefulOutputEvent">>;
 /** Additional fields recorded for each upstream Codex account attempt. */
-export type CodexAttemptLogExtra = Partial<Pick<RequestAttemptLogEntry, "errorType" | "errorMessage" | "errorCode" | "transportScope" | "retryable" | "rateLimitKind" | "cooldownReason">>;
+export type CodexAttemptLogExtra = Partial<Pick<RequestAttemptLogEntry, "errorType" | "errorMessage" | "errorCode" | "transportScope" | "retryable" | "rateLimitKind" | "cooldownReason" | "quotaResetAt" | "quotaScope">>;
 /** Minimal persistence contract needed by Codex rotating-token refreshes. */
 export type CodexRefreshTokenStore = {
     peekTokens(provider: string): Promise<StoredOAuthTokens | null>;
@@ -863,6 +874,7 @@ export type ClaudeFinalRequestLogger = (status: number, accountLabel: string, ac
     outputTokens?: number;
     cacheCreationTokens?: number;
     cacheReadTokens?: number;
+    reasoningTokens?: number;
     errorCode?: string;
     transportScope?: ProxyTransportScope;
 }) => void;
@@ -1246,15 +1258,15 @@ export type AccountQuota = {
     /** Top-level unified status. A rejected value can be authoritative even
      *  while both 5h and 7d sub-window statuses still report allowed. */
     unifiedStatus?: string;
-    /** 0.0-1.0  (from unified-5h-utilization) */
+    /** 0.0-1.0 utilization; ignore the numeric placeholder when status is unknown. */
     sessionUsed: number;
-    /** "allowed" | "throttled" | "rejected" */
+    /** "allowed" | "throttled" | "rejected" | "unknown" */
     sessionStatus: string;
     /** Unix timestamp (seconds) when the 5h window resets */
     sessionResetAt: number;
-    /** 0.0-1.0  (from unified-7d-utilization) */
+    /** 0.0-1.0 utilization; ignore the numeric placeholder when status is unknown. */
     weeklyUsed: number;
-    /** "allowed" | "throttled" | "rejected" */
+    /** "allowed" | "throttled" | "rejected" | "unknown" */
     weeklyStatus: string;
     /** Unix timestamp (seconds) when the 7d window resets */
     weeklyResetAt: number;
@@ -1700,6 +1712,8 @@ export type UpstreamAttemptContext = {
 };
 /** Token usage and rate-limit utilisation recorded at end of request. */
 export type UsageContext = {
+    /** True when input includes the cache breakdown (native Codex wire usage). */
+    inputIncludesCachedTokens?: boolean;
     inputTokens: number;
     outputTokens: number;
     cacheCreationTokens: number;
@@ -2008,6 +2022,7 @@ export type ProxyAnalysisReport = {
         requestsWithCacheRead: number;
         cacheReadTokens: number;
         cacheCreationTokens: number;
+        /** Ordinary input, excluding the separately reported cache read/write buckets. */
         inputTokens: number;
         outputTokens: number;
         requestHitRate: number | null;
@@ -2068,6 +2083,7 @@ export type ProxyAnalysisFinalRequestRecord = {
     accountType: string;
     model: string | null;
     provider: string | null;
+    inputIncludesCachedTokens?: boolean;
     inputTokens: number | null;
     outputTokens: number | null;
     cacheReadTokens: number | null;
@@ -2104,6 +2120,8 @@ export type CodexStreamUsage = {
     /** Cache writes, which bill at a premium over both reads and plain input. */
     cacheCreationTokens: number;
     reasoningTokens: number;
+    /** Whether the provider supplied a valid reasoning breakdown. */
+    reasoningTokensObserved?: boolean;
 };
 /** Semantic completion evidence observed in native Codex SSE bytes. */
 export type CodexStreamEvidence = {

package/dist/types/rag.d.ts

@@ -1,4 +1,5 @@
 import type { Tool } from "./tools.js";
+import type { EmbedInput } from "./providers.js";
 /**
  * RAG Type Definitions
  *
@@ -231,6 +232,8 @@ export type RAGPipelineConfig = {
     enableReranking?: boolean;
     /** Reranking model configuration */
     rerankingModel?: EmbeddingModelConfig;
+    /** Multi-modal RAG configuration (image + text embeddings) */
+    multiModal?: MultiModalRAGConfig;
 };
 /**
  * Ingestion options
@@ -419,6 +422,18 @@ export type VectorStore = {
         includeVectors?: boolean;
     }): Promise<VectorQueryResult[]>;
 };
+/**
+ * Structural extension of {@link VectorStore} for stores that can also write
+ * vectors via `upsert()` (e.g. {@link InMemoryVectorStore}). Custom stores
+ * that expose an `upsert` method satisfy this type without a cast.
+ */
+export type UpsertableVectorStore = VectorStore & {
+    upsert(indexName: string, items: Array<{
+        id: string;
+        vector: number[];
+        metadata?: Record<string, unknown>;
+    }>): Promise<void>;
+};
 /** Minimal shape of a query result, satisfied by both `pg` and `pglite`. */
 export type PgQueryResult<T = unknown> = {
     rows: T[];
@@ -1330,3 +1345,132 @@ export type DocumentState = {
 };
 /** Canonical RAG error code. */
 export type RAGErrorCode = "RAG_CHUNKING_ERROR" | "RAG_CHUNKING_INVALID_CONFIG" | "RAG_CHUNKING_STRATEGY_NOT_FOUND" | "RAG_CHUNKING_EMPTY_CONTENT" | "RAG_CHUNKING_SIZE_EXCEEDED" | "RAG_METADATA_EXTRACTION_ERROR" | "RAG_METADATA_EXTRACTION_TIMEOUT" | "RAG_METADATA_SCHEMA_INVALID" | "RAG_METADATA_EXTRACTOR_NOT_FOUND" | "RAG_EMBEDDING_ERROR" | "RAG_EMBEDDING_DIMENSION_MISMATCH" | "RAG_EMBEDDING_RATE_LIMIT" | "RAG_EMBEDDING_PROVIDER_ERROR" | "RAG_VECTOR_QUERY_ERROR" | "RAG_VECTOR_QUERY_TIMEOUT" | "RAG_VECTOR_STORE_UNAVAILABLE" | "RAG_VECTOR_STORE_CONNECTION_ERROR" | "RAG_VECTOR_INDEX_NOT_FOUND" | "RAG_RERANKER_ERROR" | "RAG_RERANKER_NOT_FOUND" | "RAG_RERANKER_API_ERROR" | "RAG_GRAPH_ERROR" | "RAG_GRAPH_TRAVERSAL_ERROR" | "RAG_GRAPH_NODE_NOT_FOUND" | "RAG_PIPELINE_ERROR" | "RAG_PIPELINE_STAGE_FAILED" | "RAG_PIPELINE_PARTIAL_FAILURE" | "RAG_CIRCUIT_BREAKER_OPEN" | "RAG_CIRCUIT_BREAKER_HALF_OPEN_LIMIT" | "RAG_OPERATION_TIMEOUT" | "RAG_RETRY_EXHAUSTED" | "RAG_INVALID_CONFIGURATION";
+/**
+ * Modality of an embedding model
+ */
+export type EmbeddingModality = "text" | "image" | "multimodal";
+/**
+ * Multi-modal embedding model configuration
+ */
+export type MultiModalEmbeddingConfig = {
+    /** Provider name (e.g. "bedrock") */
+    provider: string;
+    /** Model name (e.g. "amazon.titan-embed-image-v1") */
+    modelName: string;
+    /** What modalities this model supports */
+    modality: EmbeddingModality;
+    /** Embedding dimension (e.g. 1024 for Titan Image) */
+    dimensions?: number;
+};
+/**
+ * Image metadata attached to multi-modal chunks
+ */
+export type ImageChunkMetadata = {
+    /** Image width in pixels */
+    width?: number;
+    /** Image height in pixels */
+    height?: number;
+    /** Image MIME type */
+    mimeType?: string;
+    /** Image format (jpeg, png, webp, etc.) */
+    format?: string;
+    /** Original file path or URL */
+    source?: string;
+    /** Whether this chunk contains image data */
+    hasImage: boolean;
+};
+/**
+ * Extended chunk type for multi-modal RAG.
+ * Contains optional image data alongside text for image+text search.
+ */
+export type MultiModalChunk = Chunk & {
+    /** Image data as Buffer or base64 string (present for image chunks) */
+    image?: Buffer | string;
+    /** Image MIME type */
+    imageMimeType?: string;
+    /** Image-specific metadata */
+    imageMeta?: ImageChunkMetadata;
+};
+/**
+ * Multi-modal query input — can be text, image, or both.
+ * Same shape as EmbedInput.
+ */
+export type MultiModalQuery = EmbedInput;
+/**
+ * Multi-modal RAG pipeline configuration extension.
+ * Passed alongside RAGPipelineConfig to enable multi-modal features.
+ */
+export type MultiModalRAGConfig = {
+    /** Whether multi-modal RAG is enabled */
+    enabled: boolean;
+    /** Embedding model for multi-modal content */
+    embeddingModel: MultiModalEmbeddingConfig;
+    /** How to generate text from images for indexing: "caption" uses LLM vision, "filename" uses filename, "none" skips text */
+    imageTextStrategy?: "caption" | "filename" | "none";
+    /** Maximum image file size in bytes (default: 10MB) */
+    maxImageSize?: number;
+    /** Supported image MIME types */
+    supportedFormats?: string[];
+    /** Provider for generating image captions (used with imageTextStrategy: "caption") */
+    captionProvider?: string;
+    /** Model for generating image captions */
+    captionModel?: string;
+};
+/**
+ * Match type indicating how the result was found
+ */
+export type MultiModalMatchType = "text-to-text" | "text-to-image" | "image-to-image" | "image-to-text";
+/**
+ * Image search result from multi-modal retrieval
+ */
+export type MultiModalSearchResult = {
+    /** The chunk containing the match */
+    chunk: MultiModalChunk;
+    /** Similarity score (0-1) */
+    score: number;
+    /** How this match was found */
+    matchType: MultiModalMatchType;
+};
+/**
+ * Supported image MIME types for multi-modal RAG
+ */
+export declare const SUPPORTED_IMAGE_MIME_TYPES: readonly ["image/jpeg", "image/png", "image/gif", "image/webp", "image/bmp", "image/tiff", "image/avif"];
+/**
+ * Image document loaded for RAG ingestion
+ */
+export type ImageDocument = {
+    /** Unique identifier for this image document */
+    id: string;
+    /** Text representation (filename, alt text, or caption) */
+    text: string;
+    /** Raw image data */
+    image: Buffer;
+    /** Image MIME type */
+    mimeType: string;
+    /** Metadata about the image */
+    metadata: {
+        /** Original file path or URL */
+        source: string;
+        /** Image width in pixels (if detectable) */
+        width?: number;
+        /** Image height in pixels (if detectable) */
+        height?: number;
+        /** Image format (jpeg, png, webp, etc.) */
+        format?: string;
+        /** File size in bytes */
+        size: number;
+    };
+};
+/**
+ * Options for loading images
+ */
+export type ImageLoaderOptions = {
+    /** Custom metadata to attach to the image document */
+    metadata?: Record<string, unknown>;
+    /** Maximum image file size in bytes (default: 10MB) */
+    maxImageSize?: number;
+    /** Timeout for URL fetches in milliseconds (default: 30000) */
+    fetchTimeout?: number;
+    /** Custom headers for URL fetches */
+    headers?: Record<string, string>;
+};

package/dist/types/rag.js

@@ -1 +1,12 @@
-export {};
+/**
+ * Supported image MIME types for multi-modal RAG
+ */
+export const SUPPORTED_IMAGE_MIME_TYPES = [
+    "image/jpeg",
+    "image/png",
+    "image/gif",
+    "image/webp",
+    "image/bmp",
+    "image/tiff",
+    "image/avif",
+];

package/dist/utils/pricing.js

@@ -899,8 +899,10 @@ export function calculateCost(provider, model, usage) {
         // priced providers still get the cheaper cacheRead rate.
         cost += usage.cacheReadTokens * (rates.cacheRead ?? rates.input);
     }
-    if (usage.cacheCreationTokens && rates.cacheCreation) {
-        cost += usage.cacheCreationTokens * rates.cacheCreation;
+    if (usage.cacheCreationTokens) {
+        // Like cache reads, disjoint cache writes must not silently become free
+        // when a model has no separate creation rate. Preserve explicit zero rates.
+        cost += usage.cacheCreationTokens * (rates.cacheCreation ?? rates.input);
     }
     return Math.round(cost * 1_000_000) / 1_000_000; // Round to 6 decimal places
 }