@juspay/neurolink 8.5.0 → 8.6.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (58) hide show
  1. package/CHANGELOG.md +17 -0
  2. package/dist/adapters/providerImageAdapter.d.ts +4 -2
  3. package/dist/adapters/providerImageAdapter.js +72 -11
  4. package/dist/config/conversationMemory.d.ts +6 -0
  5. package/dist/config/conversationMemory.js +14 -0
  6. package/dist/constants/enums.d.ts +23 -3
  7. package/dist/constants/enums.js +30 -4
  8. package/dist/constants/tokens.d.ts +27 -12
  9. package/dist/constants/tokens.js +46 -12
  10. package/dist/core/baseProvider.js +6 -2
  11. package/dist/core/modules/GenerationHandler.js +20 -5
  12. package/dist/core/modules/MessageBuilder.js +4 -0
  13. package/dist/core/modules/TelemetryHandler.js +6 -1
  14. package/dist/lib/adapters/providerImageAdapter.d.ts +4 -2
  15. package/dist/lib/adapters/providerImageAdapter.js +72 -11
  16. package/dist/lib/config/conversationMemory.d.ts +6 -0
  17. package/dist/lib/config/conversationMemory.js +14 -0
  18. package/dist/lib/constants/enums.d.ts +23 -3
  19. package/dist/lib/constants/enums.js +30 -4
  20. package/dist/lib/constants/tokens.d.ts +27 -12
  21. package/dist/lib/constants/tokens.js +46 -12
  22. package/dist/lib/core/baseProvider.js +6 -2
  23. package/dist/lib/core/modules/GenerationHandler.js +20 -5
  24. package/dist/lib/core/modules/MessageBuilder.js +4 -0
  25. package/dist/lib/core/modules/TelemetryHandler.js +6 -1
  26. package/dist/lib/middleware/builtin/guardrails.js +7 -0
  27. package/dist/lib/models/modelRegistry.js +93 -0
  28. package/dist/lib/neurolink.js +75 -5
  29. package/dist/lib/providers/googleAiStudio.d.ts +27 -0
  30. package/dist/lib/providers/googleAiStudio.js +27 -0
  31. package/dist/lib/providers/googleVertex.d.ts +35 -0
  32. package/dist/lib/providers/googleVertex.js +38 -0
  33. package/dist/lib/telemetry/telemetryService.d.ts +1 -1
  34. package/dist/lib/telemetry/telemetryService.js +4 -4
  35. package/dist/lib/types/common.d.ts +5 -0
  36. package/dist/lib/types/content.d.ts +1 -1
  37. package/dist/lib/types/generateTypes.d.ts +68 -2
  38. package/dist/lib/types/multimodal.d.ts +38 -1
  39. package/dist/lib/types/streamTypes.d.ts +21 -2
  40. package/dist/lib/utils/messageBuilder.js +70 -8
  41. package/dist/lib/utils/multimodalOptionsBuilder.d.ts +1 -1
  42. package/dist/middleware/builtin/guardrails.js +7 -0
  43. package/dist/models/modelRegistry.js +93 -0
  44. package/dist/neurolink.js +75 -5
  45. package/dist/providers/googleAiStudio.d.ts +27 -0
  46. package/dist/providers/googleAiStudio.js +27 -0
  47. package/dist/providers/googleVertex.d.ts +35 -0
  48. package/dist/providers/googleVertex.js +38 -0
  49. package/dist/telemetry/telemetryService.d.ts +1 -1
  50. package/dist/telemetry/telemetryService.js +4 -4
  51. package/dist/types/common.d.ts +5 -0
  52. package/dist/types/content.d.ts +1 -1
  53. package/dist/types/generateTypes.d.ts +68 -2
  54. package/dist/types/multimodal.d.ts +38 -1
  55. package/dist/types/streamTypes.d.ts +21 -2
  56. package/dist/utils/messageBuilder.js +70 -8
  57. package/dist/utils/multimodalOptionsBuilder.d.ts +1 -1
  58. package/package.json +1 -1
@@ -1,7 +1,7 @@
1
1
  import type { Tool } from "ai";
2
2
  import type { ValidationSchema, StandardRecord } from "./typeAliases.js";
3
3
  import type { AIModelProviderConfig } from "./providers.js";
4
- import type { Content } from "./content.js";
4
+ import type { Content, ImageWithAltText } from "./content.js";
5
5
  import type { AnalyticsData, ToolExecutionEvent, ToolExecutionSummary } from "../types/index.js";
6
6
  import { AIProviderName } from "../constants/enums.js";
7
7
  import type { TokenUsage } from "./analytics.js";
@@ -125,7 +125,24 @@ export type StreamOptions = {
125
125
  input: {
126
126
  text: string;
127
127
  audio?: AudioInputSpec;
128
- images?: Array<Buffer | string>;
128
+ /**
129
+ * Images to include in the request.
130
+ * Supports simple image data (Buffer, string) or objects with alt text for accessibility.
131
+ *
132
+ * @example Simple usage
133
+ * ```typescript
134
+ * images: [imageBuffer, "https://example.com/image.jpg"]
135
+ * ```
136
+ *
137
+ * @example With alt text for accessibility
138
+ * ```typescript
139
+ * images: [
140
+ * { data: imageBuffer, altText: "Product screenshot showing main dashboard" },
141
+ * { data: "https://example.com/chart.png", altText: "Sales chart for Q3 2024" }
142
+ * ]
143
+ * ```
144
+ */
145
+ images?: Array<Buffer | string | ImageWithAltText>;
129
146
  csvFiles?: Array<Buffer | string>;
130
147
  pdfFiles?: Array<Buffer | string>;
131
148
  files?: Array<Buffer | string>;
@@ -211,6 +228,8 @@ export type StreamResult = {
211
228
  totalToolExecutions?: number;
212
229
  toolExecutionTime?: number;
213
230
  hasToolErrors?: boolean;
231
+ guardrailsBlocked?: boolean;
232
+ error?: string;
214
233
  };
215
234
  analytics?: AnalyticsData | Promise<AnalyticsData>;
216
235
  evaluation?: EvaluationData | Promise<EvaluationData>;
@@ -3,13 +3,37 @@
3
3
  * Centralized logic for building message arrays from TextGenerationOptions
4
4
  * Enhanced with multimodal support for images
5
5
  */
6
- import { CONVERSATION_INSTRUCTIONS } from "../config/conversationMemory.js";
6
+ import { CONVERSATION_INSTRUCTIONS, STRUCTURED_OUTPUT_INSTRUCTIONS, } from "../config/conversationMemory.js";
7
7
  import { ProviderImageAdapter, MultimodalLogger, } from "../adapters/providerImageAdapter.js";
8
8
  import { logger } from "./logger.js";
9
9
  import { FileDetector } from "./fileDetector.js";
10
10
  import { PDFProcessor } from "./pdfProcessor.js";
11
11
  import { request, getGlobalDispatcher, interceptors } from "undici";
12
12
  import { readFileSync, existsSync } from "fs";
13
+ /**
14
+ * Type guard to check if an image input has alt text
15
+ */
16
+ function isImageWithAltText(image) {
17
+ return (typeof image === "object" && !Buffer.isBuffer(image) && "data" in image);
18
+ }
19
+ /**
20
+ * Extract image data from an image input (handles both simple and alt text formats)
21
+ */
22
+ function extractImageData(image) {
23
+ if (isImageWithAltText(image)) {
24
+ return image.data;
25
+ }
26
+ return image;
27
+ }
28
+ /**
29
+ * Extract alt text from an image input if available
30
+ */
31
+ function extractAltText(image) {
32
+ if (isImageWithAltText(image)) {
33
+ return image.altText;
34
+ }
35
+ return undefined;
36
+ }
13
37
  /**
14
38
  * Type guard for validating message roles
15
39
  */
@@ -199,6 +223,15 @@ function formatCSVMetadata(metadata) {
199
223
  }
200
224
  return parts.length > 0 ? `**Metadata**: ${parts.join(" | ")}` : "";
201
225
  }
226
+ /**
227
+ * Check if structured output mode should be enabled
228
+ * Structured output is used when a schema is provided with json/structured format
229
+ */
230
+ function shouldUseStructuredOutput(options) {
231
+ return (!!options.schema &&
232
+ (options.output?.format === "json" ||
233
+ options.output?.format === "structured"));
234
+ }
202
235
  /**
203
236
  * Build a properly formatted message array for AI providers
204
237
  * Combines system prompt, conversation history, and current user prompt
@@ -215,6 +248,10 @@ export async function buildMessagesArray(options) {
215
248
  if (hasConversationHistory) {
216
249
  systemPrompt = `${systemPrompt.trim()}${CONVERSATION_INSTRUCTIONS}`;
217
250
  }
251
+ // Add structured output instructions when schema is provided with json/structured format
252
+ if (shouldUseStructuredOutput(options)) {
253
+ systemPrompt = `${systemPrompt.trim()}${STRUCTURED_OUTPUT_INSTRUCTIONS}`;
254
+ }
218
255
  // Add system message if we have one
219
256
  if (systemPrompt.trim()) {
220
257
  messages.push({
@@ -473,6 +510,10 @@ export async function buildMultimodalMessagesArray(options, provider, model) {
473
510
  if (hasConversationHistory) {
474
511
  systemPrompt = `${systemPrompt.trim()}${CONVERSATION_INSTRUCTIONS}`;
475
512
  }
513
+ // Add structured output instructions when schema is provided with json/structured format
514
+ if (shouldUseStructuredOutput(options)) {
515
+ systemPrompt = `${systemPrompt.trim()}${STRUCTURED_OUTPUT_INSTRUCTIONS}`;
516
+ }
476
517
  // Add file handling guidance when multimodal files are present
477
518
  const hasCSVFiles = (options.input.csvFiles && options.input.csvFiles.length > 0) ||
478
519
  (options.input.files &&
@@ -622,28 +663,47 @@ async function downloadImageFromUrl(url) {
622
663
  * - URLs: Downloaded and converted to base64 for Vercel AI SDK compatibility
623
664
  * - Local files: Converted to base64 for Vercel AI SDK compatibility
624
665
  * - Buffers/Data URIs: Processed normally
666
+ * - Supports alt text for accessibility (included as context in text parts)
625
667
  */
626
668
  async function convertSimpleImagesToProviderFormat(text, images, provider, _model) {
627
669
  // For Vercel AI SDK, we need to return the content in the standard format
628
670
  // The Vercel AI SDK will handle provider-specific formatting internally
671
+ // IMPORTANT: Generate alt text descriptions BEFORE URL downloading to maintain correct image numbering
672
+ // This ensures image numbers match the original order provided by users, even if some URLs fail to download
673
+ const altTextDescriptions = images
674
+ .map((image, idx) => {
675
+ const altText = extractAltText(image);
676
+ return altText ? `[Image ${idx + 1}: ${altText}]` : null;
677
+ })
678
+ .filter(Boolean);
679
+ // Build enhanced text with alt text context for accessibility
680
+ // NOTE: Alt text is appended to the user's prompt as contextual information because most AI providers
681
+ // don't have native alt text fields in their APIs. This approach ensures accessibility metadata
682
+ // is preserved and helps AI models better understand image content.
683
+ const enhancedText = altTextDescriptions.length > 0
684
+ ? `${text}\n\nImage descriptions for context: ${altTextDescriptions.join(" ")}`
685
+ : text;
629
686
  // Smart auto-detection: separate URLs from actual image data
687
+ // Also track alt text for each image
630
688
  const urlImages = [];
631
689
  const actualImages = [];
632
690
  images.forEach((image, _index) => {
633
- if (typeof image === "string" && isInternetUrl(image)) {
691
+ const imageData = extractImageData(image);
692
+ const altText = extractAltText(image);
693
+ if (typeof imageData === "string" && isInternetUrl(imageData)) {
634
694
  // Internet URL - will be downloaded and converted to base64
635
- urlImages.push(image);
695
+ urlImages.push({ url: imageData, altText });
636
696
  }
637
697
  else {
638
698
  // Actual image data (file path, Buffer, data URI) - process for Vercel AI SDK
639
- actualImages.push(image);
699
+ actualImages.push({ data: imageData, altText });
640
700
  }
641
701
  });
642
702
  // Download URL images and add to actual images
643
- for (const url of urlImages) {
703
+ for (const { url, altText } of urlImages) {
644
704
  try {
645
705
  const downloadedDataUri = await downloadImageFromUrl(url);
646
- actualImages.push(downloadedDataUri);
706
+ actualImages.push({ data: downloadedDataUri, altText });
647
707
  }
648
708
  catch (error) {
649
709
  MultimodalLogger.logError("URL_DOWNLOAD_FAILED_SKIPPING", error, { url });
@@ -651,9 +711,11 @@ async function convertSimpleImagesToProviderFormat(text, images, provider, _mode
651
711
  logger.warn(`Failed to download image from ${url}, skipping: ${error instanceof Error ? error.message : String(error)}`);
652
712
  }
653
713
  }
654
- const content = [{ type: "text", text }];
714
+ const content = [
715
+ { type: "text", text: enhancedText },
716
+ ];
655
717
  // Process all images (including downloaded URLs) for Vercel AI SDK
656
- actualImages.forEach((image, index) => {
718
+ actualImages.forEach(({ data: image }, index) => {
657
719
  try {
658
720
  // Vercel AI SDK expects { type: 'image', image: Buffer | string, mimeType?: string }
659
721
  // For Vertex AI, we need to include mimeType
@@ -44,7 +44,7 @@ import type { StreamOptions } from "../types/streamTypes.js";
44
44
  export declare function buildMultimodalOptions(options: StreamOptions, providerName: string, modelName: string): {
45
45
  input: {
46
46
  text: string;
47
- images: (string | Buffer<ArrayBufferLike>)[] | undefined;
47
+ images: (string | Buffer<ArrayBufferLike> | import("../types/multimodal.js").ImageWithAltText)[] | undefined;
48
48
  content: import("../types/multimodal.js").Content[] | undefined;
49
49
  files: (string | Buffer<ArrayBufferLike>)[] | undefined;
50
50
  csvFiles: (string | Buffer<ArrayBufferLike>)[] | undefined;
@@ -69,8 +69,10 @@ export function createGuardrailsMiddleware(config = {}) {
69
69
  };
70
70
  }
71
71
  const { stream, ...rest } = await doStream();
72
+ let hasYieldedChunks = false;
72
73
  const transformStream = new TransformStream({
73
74
  transform(chunk, controller) {
75
+ hasYieldedChunks = true;
74
76
  let filteredChunk = chunk;
75
77
  if (typeof filteredChunk === "object" &&
76
78
  "textDelta" in filteredChunk) {
@@ -84,6 +86,11 @@ export function createGuardrailsMiddleware(config = {}) {
84
86
  }
85
87
  controller.enqueue(filteredChunk);
86
88
  },
89
+ flush() {
90
+ if (!hasYieldedChunks) {
91
+ logger.warn(`[GuardrailsMiddleware] Stream ended without yielding any chunks`);
92
+ }
93
+ },
87
94
  });
88
95
  return {
89
96
  stream: stream.pipeThrough(transformStream),
@@ -188,6 +188,99 @@ export const MODEL_REGISTRY = {
188
188
  category: "general",
189
189
  },
190
190
  // Anthropic Models
191
+ [AnthropicModels.CLAUDE_OPUS_4_5]: {
192
+ id: AnthropicModels.CLAUDE_OPUS_4_5,
193
+ name: "Claude Opus 4.5",
194
+ provider: AIProviderName.ANTHROPIC,
195
+ description: "Anthropic's most capable model with exceptional reasoning, coding, and multimodal capabilities",
196
+ capabilities: {
197
+ vision: true,
198
+ functionCalling: true,
199
+ codeGeneration: true,
200
+ reasoning: true,
201
+ multimodal: true,
202
+ streaming: true,
203
+ jsonMode: false,
204
+ },
205
+ pricing: {
206
+ inputCostPer1K: 0.015,
207
+ outputCostPer1K: 0.075,
208
+ currency: "USD",
209
+ },
210
+ performance: {
211
+ speed: "medium",
212
+ quality: "high",
213
+ accuracy: "high",
214
+ },
215
+ limits: {
216
+ maxContextTokens: 200000,
217
+ maxOutputTokens: 64000,
218
+ maxRequestsPerMinute: 50,
219
+ },
220
+ useCases: {
221
+ coding: 10,
222
+ creative: 10,
223
+ analysis: 10,
224
+ conversation: 9,
225
+ reasoning: 10,
226
+ translation: 9,
227
+ summarization: 9,
228
+ },
229
+ aliases: [
230
+ "claude-4.5-opus",
231
+ "claude-opus-latest",
232
+ "opus-4.5",
233
+ "anthropic-flagship",
234
+ ],
235
+ deprecated: false,
236
+ isLocal: false,
237
+ releaseDate: "2025-11-24",
238
+ category: "reasoning",
239
+ },
240
+ [AnthropicModels.CLAUDE_SONNET_4_5]: {
241
+ id: AnthropicModels.CLAUDE_SONNET_4_5,
242
+ name: "Claude Sonnet 4.5",
243
+ provider: AIProviderName.ANTHROPIC,
244
+ description: "Balanced Claude model with excellent performance across all tasks including vision and reasoning",
245
+ capabilities: {
246
+ vision: true,
247
+ functionCalling: true,
248
+ codeGeneration: true,
249
+ reasoning: true,
250
+ multimodal: true,
251
+ streaming: true,
252
+ jsonMode: false,
253
+ },
254
+ pricing: {
255
+ inputCostPer1K: 0.003,
256
+ outputCostPer1K: 0.015,
257
+ currency: "USD",
258
+ },
259
+ performance: {
260
+ speed: "medium",
261
+ quality: "high",
262
+ accuracy: "high",
263
+ },
264
+ limits: {
265
+ maxContextTokens: 200000,
266
+ maxOutputTokens: 64000,
267
+ maxRequestsPerMinute: 100,
268
+ },
269
+ useCases: {
270
+ coding: 10,
271
+ creative: 9,
272
+ analysis: 9,
273
+ conversation: 9,
274
+ reasoning: 10,
275
+ translation: 8,
276
+ summarization: 8,
277
+ },
278
+ aliases: ["claude-4.5-sonnet", "claude-sonnet-latest", "sonnet-4.5"],
279
+ deprecated: false,
280
+ isLocal: false,
281
+ releaseDate: "2025-09-29",
282
+ category: "coding",
283
+ },
191
284
  [AnthropicModels.CLAUDE_4_5_HAIKU]: {
192
285
  id: AnthropicModels.CLAUDE_4_5_HAIKU,
193
286
  name: "Claude 4.5 Haiku",
package/dist/neurolink.js CHANGED
@@ -1998,19 +1998,85 @@ Current user's request: ${currentInput}`;
1998
1998
  }
1999
1999
  }
2000
2000
  const { stream: mcpStream, provider: providerName } = await this.createMCPStream(enhancedOptions);
2001
- // Create a wrapper around the stream that accumulates content
2002
2001
  let accumulatedContent = "";
2002
+ let chunkCount = 0;
2003
+ const metadata = {
2004
+ fallbackAttempted: false,
2005
+ guardrailsBlocked: false,
2006
+ error: undefined,
2007
+ };
2003
2008
  const processedStream = (async function* (self) {
2004
2009
  try {
2005
2010
  for await (const chunk of mcpStream) {
2011
+ chunkCount++;
2006
2012
  if (chunk &&
2007
2013
  "content" in chunk &&
2008
2014
  typeof chunk.content === "string") {
2009
2015
  accumulatedContent += chunk.content;
2010
- // Emit chunk event for compatibility
2011
2016
  self.emitter.emit("response:chunk", chunk.content);
2012
2017
  }
2013
- yield chunk; // Preserve original streaming behavior
2018
+ yield chunk;
2019
+ }
2020
+ if (chunkCount === 0 && !metadata.fallbackAttempted) {
2021
+ metadata.fallbackAttempted = true;
2022
+ const errorMsg = "Stream completed with 0 chunks (possible guardrails block)";
2023
+ metadata.error = errorMsg;
2024
+ const fallbackRoute = ModelRouter.getFallbackRoute(originalPrompt || enhancedOptions.input.text || "", {
2025
+ provider: providerName,
2026
+ model: enhancedOptions.model || "gpt-4o",
2027
+ reasoning: "primary failed",
2028
+ confidence: 0.5,
2029
+ }, { fallbackStrategy: "auto" });
2030
+ logger.warn("Retrying with fallback provider", {
2031
+ originalProvider: providerName,
2032
+ fallbackProvider: fallbackRoute.provider,
2033
+ reason: errorMsg,
2034
+ });
2035
+ try {
2036
+ const fallbackProvider = await AIProviderFactory.createProvider(fallbackRoute.provider, fallbackRoute.model);
2037
+ // Ensure fallback provider can execute tools
2038
+ fallbackProvider.setupToolExecutor({
2039
+ customTools: self.getCustomTools(),
2040
+ executeTool: self.executeTool.bind(self),
2041
+ }, "NeuroLink.fallbackStream");
2042
+ // Get conversation messages for context (same as primary stream)
2043
+ const conversationMessages = await getConversationMessages(self.conversationMemory, {
2044
+ prompt: enhancedOptions.input.text,
2045
+ context: enhancedOptions.context,
2046
+ });
2047
+ const fallbackResult = await fallbackProvider.stream({
2048
+ ...enhancedOptions,
2049
+ model: fallbackRoute.model,
2050
+ conversationMessages,
2051
+ });
2052
+ let fallbackChunkCount = 0;
2053
+ for await (const fallbackChunk of fallbackResult.stream) {
2054
+ fallbackChunkCount++;
2055
+ if (fallbackChunk &&
2056
+ "content" in fallbackChunk &&
2057
+ typeof fallbackChunk.content === "string") {
2058
+ accumulatedContent += fallbackChunk.content;
2059
+ self.emitter.emit("response:chunk", fallbackChunk.content);
2060
+ }
2061
+ yield fallbackChunk;
2062
+ }
2063
+ if (fallbackChunkCount === 0) {
2064
+ throw new Error(`Fallback provider ${fallbackRoute.provider} also returned 0 chunks`);
2065
+ }
2066
+ // Fallback succeeded - likely guardrails blocked primary
2067
+ metadata.guardrailsBlocked = true;
2068
+ }
2069
+ catch (fallbackError) {
2070
+ const fallbackErrorMsg = fallbackError instanceof Error
2071
+ ? fallbackError.message
2072
+ : String(fallbackError);
2073
+ metadata.error = `${errorMsg}; Fallback failed: ${fallbackErrorMsg}`;
2074
+ logger.error("Fallback provider failed", {
2075
+ fallbackProvider: fallbackRoute.provider,
2076
+ error: fallbackErrorMsg,
2077
+ });
2078
+ throw fallbackError;
2079
+ }
2014
2080
  }
2015
2081
  }
2016
2082
  finally {
@@ -2053,7 +2119,7 @@ Current user's request: ${currentInput}`;
2053
2119
  }
2054
2120
  }
2055
2121
  })(this);
2056
- const streamResult = await this.processStreamResult(mcpStream, enhancedOptions, factoryResult);
2122
+ const streamResult = await this.processStreamResult(processedStream, enhancedOptions, factoryResult);
2057
2123
  const responseTime = Date.now() - startTime;
2058
2124
  this.emitStreamEndEvents(streamResult);
2059
2125
  return this.createStreamResponse(streamResult, processedStream, {
@@ -2062,7 +2128,9 @@ Current user's request: ${currentInput}`;
2062
2128
  startTime,
2063
2129
  responseTime,
2064
2130
  streamId,
2065
- fallback: false,
2131
+ fallback: metadata.fallbackAttempted,
2132
+ guardrailsBlocked: metadata.guardrailsBlocked,
2133
+ error: metadata.error,
2066
2134
  });
2067
2135
  }
2068
2136
  catch (error) {
@@ -2181,6 +2249,8 @@ Current user's request: ${currentInput}`;
2181
2249
  startTime: config.startTime,
2182
2250
  responseTime: config.responseTime,
2183
2251
  fallback: config.fallback || false,
2252
+ guardrailsBlocked: config.guardrailsBlocked,
2253
+ error: config.error,
2184
2254
  },
2185
2255
  };
2186
2256
  }
@@ -6,6 +6,33 @@ import { BaseProvider } from "../core/baseProvider.js";
6
6
  /**
7
7
  * Google AI Studio provider implementation using BaseProvider
8
8
  * Migrated from original GoogleAIStudio class to new factory pattern
9
+ *
10
+ * @important Structured Output Limitation
11
+ * Google Gemini models cannot combine function calling (tools) with structured
12
+ * output (JSON schema). When using schemas with output.format: "json", you MUST
13
+ * set disableTools: true.
14
+ *
15
+ * Error without disableTools:
16
+ * "Function calling with a response mime type: 'application/json' is unsupported"
17
+ *
18
+ * This is a Google API limitation documented at:
19
+ * https://ai.google.dev/gemini-api/docs/function-calling
20
+ *
21
+ * @example
22
+ * ```typescript
23
+ * // ✅ Correct usage with schemas
24
+ * const provider = new GoogleAIStudioProvider("gemini-2.5-flash");
25
+ * const result = await provider.generate({
26
+ * input: { text: "Analyze data" },
27
+ * schema: MySchema,
28
+ * output: { format: "json" },
29
+ * disableTools: true // Required
30
+ * });
31
+ * ```
32
+ *
33
+ * @note Gemini 3 Pro Preview (November 2025) will support combining tools + schemas
34
+ * @note "Too many states for serving" errors can occur with complex schemas + tools.
35
+ * Solution: Simplify schema or use disableTools: true
9
36
  */
10
37
  export declare class GoogleAIStudioProvider extends BaseProvider {
11
38
  constructor(modelName?: string, sdk?: unknown);
@@ -27,6 +27,33 @@ if (!process.env.GOOGLE_GENERATIVE_AI_API_KEY &&
27
27
  /**
28
28
  * Google AI Studio provider implementation using BaseProvider
29
29
  * Migrated from original GoogleAIStudio class to new factory pattern
30
+ *
31
+ * @important Structured Output Limitation
32
+ * Google Gemini models cannot combine function calling (tools) with structured
33
+ * output (JSON schema). When using schemas with output.format: "json", you MUST
34
+ * set disableTools: true.
35
+ *
36
+ * Error without disableTools:
37
+ * "Function calling with a response mime type: 'application/json' is unsupported"
38
+ *
39
+ * This is a Google API limitation documented at:
40
+ * https://ai.google.dev/gemini-api/docs/function-calling
41
+ *
42
+ * @example
43
+ * ```typescript
44
+ * // ✅ Correct usage with schemas
45
+ * const provider = new GoogleAIStudioProvider("gemini-2.5-flash");
46
+ * const result = await provider.generate({
47
+ * input: { text: "Analyze data" },
48
+ * schema: MySchema,
49
+ * output: { format: "json" },
50
+ * disableTools: true // Required
51
+ * });
52
+ * ```
53
+ *
54
+ * @note Gemini 3 Pro Preview (November 2025) will support combining tools + schemas
55
+ * @note "Too many states for serving" errors can occur with complex schemas + tools.
56
+ * Solution: Simplify schema or use disableTools: true
30
57
  */
31
58
  export class GoogleAIStudioProvider extends BaseProvider {
32
59
  constructor(modelName, sdk) {
@@ -13,6 +13,41 @@ import { BaseProvider } from "../core/baseProvider.js";
13
13
  * - Fresh model creation for each request
14
14
  * - Enhanced error handling with setup guidance
15
15
  * - Tool registration and context management
16
+ *
17
+ * @important Structured Output Limitation (Gemini Models Only)
18
+ * Google Gemini models on Vertex AI cannot combine function calling (tools) with
19
+ * structured output (JSON schema). When using schemas, you MUST set disableTools: true.
20
+ *
21
+ * Error without disableTools:
22
+ * "Function calling with a response mime type: 'application/json' is unsupported"
23
+ *
24
+ * This limitation ONLY affects Gemini models. Anthropic Claude models via Vertex
25
+ * AI do NOT have this limitation and support both tools + schemas simultaneously.
26
+ *
27
+ * @example Gemini models with schemas
28
+ * ```typescript
29
+ * const provider = new GoogleVertexProvider("gemini-2.5-flash");
30
+ * const result = await provider.generate({
31
+ * input: { text: "Analyze data" },
32
+ * schema: MySchema,
33
+ * output: { format: "json" },
34
+ * disableTools: true // Required for Gemini models
35
+ * });
36
+ * ```
37
+ *
38
+ * @example Claude models (no limitation)
39
+ * ```typescript
40
+ * const provider = new GoogleVertexProvider("claude-3-5-sonnet-20241022");
41
+ * const result = await provider.generate({
42
+ * input: { text: "Analyze data" },
43
+ * schema: MySchema,
44
+ * output: { format: "json" }
45
+ * // No disableTools needed - Claude supports both
46
+ * });
47
+ * ```
48
+ *
49
+ * @note Gemini 3 Pro Preview (November 2025) will support combining tools + schemas
50
+ * @see https://cloud.google.com/vertex-ai/docs/generative-ai/learn/models
16
51
  */
17
52
  export declare class GoogleVertexProvider extends BaseProvider {
18
53
  private projectId;
@@ -234,6 +234,41 @@ const isAnthropicModel = (modelName) => {
234
234
  * - Fresh model creation for each request
235
235
  * - Enhanced error handling with setup guidance
236
236
  * - Tool registration and context management
237
+ *
238
+ * @important Structured Output Limitation (Gemini Models Only)
239
+ * Google Gemini models on Vertex AI cannot combine function calling (tools) with
240
+ * structured output (JSON schema). When using schemas, you MUST set disableTools: true.
241
+ *
242
+ * Error without disableTools:
243
+ * "Function calling with a response mime type: 'application/json' is unsupported"
244
+ *
245
+ * This limitation ONLY affects Gemini models. Anthropic Claude models via Vertex
246
+ * AI do NOT have this limitation and support both tools + schemas simultaneously.
247
+ *
248
+ * @example Gemini models with schemas
249
+ * ```typescript
250
+ * const provider = new GoogleVertexProvider("gemini-2.5-flash");
251
+ * const result = await provider.generate({
252
+ * input: { text: "Analyze data" },
253
+ * schema: MySchema,
254
+ * output: { format: "json" },
255
+ * disableTools: true // Required for Gemini models
256
+ * });
257
+ * ```
258
+ *
259
+ * @example Claude models (no limitation)
260
+ * ```typescript
261
+ * const provider = new GoogleVertexProvider("claude-3-5-sonnet-20241022");
262
+ * const result = await provider.generate({
263
+ * input: { text: "Analyze data" },
264
+ * schema: MySchema,
265
+ * output: { format: "json" }
266
+ * // No disableTools needed - Claude supports both
267
+ * });
268
+ * ```
269
+ *
270
+ * @note Gemini 3 Pro Preview (November 2025) will support combining tools + schemas
271
+ * @see https://cloud.google.com/vertex-ai/docs/generative-ai/learn/models
237
272
  */
238
273
  export class GoogleVertexProvider extends BaseProvider {
239
274
  projectId;
@@ -1363,11 +1398,14 @@ export class GoogleVertexProvider extends BaseProvider {
1363
1398
  getModelSuggestions(requestedModel) {
1364
1399
  const availableModels = {
1365
1400
  google: [
1401
+ "gemini-3-pro-preview-11-2025",
1402
+ "gemini-3-pro-latest",
1366
1403
  "gemini-3-pro-preview",
1367
1404
  "gemini-2.5-pro",
1368
1405
  "gemini-2.5-flash",
1369
1406
  "gemini-2.5-flash-lite",
1370
1407
  "gemini-2.0-flash-001",
1408
+ "gemini-2.0-flash-lite",
1371
1409
  "gemini-1.5-pro",
1372
1410
  "gemini-1.5-flash",
1373
1411
  ],
@@ -31,7 +31,7 @@ export declare class TelemetryService {
31
31
  private initializeTelemetry;
32
32
  private initializeMetrics;
33
33
  initialize(): Promise<void>;
34
- traceAIRequest<T>(provider: string, operation: () => Promise<T>): Promise<T>;
34
+ traceAIRequest<T>(provider: string, operation: () => Promise<T>, operationType?: string): Promise<T>;
35
35
  recordAIRequest(provider: string, model: string, tokens: number, duration: number): void;
36
36
  recordAIError(provider: string, error: Error): void;
37
37
  recordMCPToolCall(toolName: string, duration: number, success: boolean): void;
@@ -108,14 +108,14 @@ export class TelemetryService {
108
108
  }
109
109
  }
110
110
  // AI Operation Tracing (NO-OP when disabled)
111
- async traceAIRequest(provider, operation) {
111
+ async traceAIRequest(provider, operation, operationType = "generate_text") {
112
112
  if (!this.enabled || !this.tracer) {
113
- return await operation(); // Direct execution when disabled
113
+ return await operation();
114
114
  }
115
- const span = this.tracer.startSpan(`ai.${provider}.generate_text`, {
115
+ const span = this.tracer.startSpan(`ai.${provider}.${operationType}`, {
116
116
  attributes: {
117
117
  "ai.provider": provider,
118
- "ai.operation": "generate_text",
118
+ "ai.operation": operationType,
119
119
  },
120
120
  });
121
121
  try {
@@ -129,3 +129,8 @@ export type TypedEventEmitter<TEvents extends Record<string, unknown>> = {
129
129
  listenerCount<K extends keyof TEvents>(event: K): number;
130
130
  listeners<K extends keyof TEvents>(event: K): Array<(...args: unknown[]) => void>;
131
131
  };
132
+ export type Context = {
133
+ traceName?: string;
134
+ userId?: string;
135
+ sessionId?: string;
136
+ };
@@ -14,5 +14,5 @@
14
14
  * import type { MultimodalInput } from './types/multimodal.js';
15
15
  * ```
16
16
  */
17
- export type { TextContent, ImageContent, CSVContent, PDFContent, AudioContent, VideoContent, Content, MultimodalInput, MultimodalMessage, VisionCapability, ProviderImageFormat, ProcessedImage, ProviderMultimodalPayload, } from "./multimodal.js";
17
+ export type { TextContent, ImageContent, CSVContent, PDFContent, AudioContent, VideoContent, Content, ImageWithAltText, MultimodalInput, MultimodalMessage, VisionCapability, ProviderImageFormat, ProcessedImage, ProviderMultimodalPayload, } from "./multimodal.js";
18
18
  export { isTextContent, isImageContent, isCSVContent, isPDFContent, isAudioContent, isVideoContent, isMultimodalInput, } from "./multimodal.js";