@juspay/neurolink 9.68.14 → 9.68.16

package/dist/lib/providers/cohere.d.ts

@@ -1,27 +1,26 @@
 import type { AIProviderName } from "../constants/enums.js";
-import { BaseProvider } from "../core/baseProvider.js";
-import type { NeurolinkCredentials, StreamOptions, StreamResult, ValidationSchema } from "../types/index.js";
-import type { LanguageModel } from "../types/index.js";
+import type { NeurolinkCredentials } from "../types/index.js";
+import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
 /**
- * Cohere Provider
+ * Cohere Provider — direct HTTP, no AI SDK.
  *
- * Routes Command R / Command R+ chat completions through Cohere's OpenAI-
- * compatible endpoint. Embed v3 and Rerank v3 are top-tier for RAG but are
- * accessed via the Cohere native SDK / dedicated embedding routes (out of
- * scope for the LLM provider).
+ * Routes Command R / Command R+ chat completions through Cohere's
+ * OpenAI-compatible endpoint at /compatibility/v1. All request/stream/
+ * tool-loop orchestration lives in `OpenAIChatCompletionsProvider`; this
+ * class only declares configuration and provider-specific error mapping.
+ *
+ * Embed v3 is exposed via `embed()` / `embedMany()` backed by a native
+ * POST to /v2/embed (the compatibility path is chat-only).
  *
  * @see https://docs.cohere.com/docs/compatibility-api
+ * @see https://docs.cohere.com/reference/embed
  */
-export declare class CohereProvider extends BaseProvider {
-    private model;
-    private apiKey;
-    private baseURL;
+export declare class CohereProvider extends OpenAIChatCompletionsProvider {
     constructor(modelName?: string, sdk?: unknown, _region?: string, credentials?: NeurolinkCredentials["cohere"]);
-    protected executeStream(options: StreamOptions, _analysisSchema?: ValidationSchema): Promise<StreamResult>;
-    private executeStreamInner;
     protected getProviderName(): AIProviderName;
     protected getDefaultModel(): string;
-    protected getAISDKModel(): LanguageModel;
+    protected getFallbackModelName(): string;
+    protected getFallbackModels(): string[];
     protected formatProviderError(error: unknown): Error;
     validateConfiguration(): Promise<boolean>;
     getConfiguration(): {
@@ -46,7 +45,10 @@ export declare class CohereProvider extends BaseProvider {
     /**
      * Batch embedding via Cohere's native /v2/embed endpoint. Cohere caps at
      * 96 inputs per request; larger batches are chunked.
+     *
+     * Partial failures are not surfaced: if any chunk request fails the whole
+     * call rejects and already-embedded chunks are discarded — callers should
+     * retry the full input.
      */
     embedMany(texts: string[], modelName?: string): Promise<number[][]>;
 }
-export default CohereProvider;

package/dist/lib/providers/cohere.js

@@ -1,154 +1,61 @@
-import { createOpenAI } from "@ai-sdk/openai";
 import { CohereModels } from "../constants/enums.js";
-import { BaseProvider } from "../core/baseProvider.js";
-import { DEFAULT_MAX_STEPS } from "../core/constants.js";
-import { streamAnalyticsCollector } from "../core/streamAnalytics.js";
-import { isNeuroLink } from "../neurolink.js";
-import { createLoggingFetch } from "../utils/loggingFetch.js";
-import { tracers, ATTR, withClientStreamSpan } from "../telemetry/index.js";
 import { AuthenticationError, InvalidModelError, NetworkError, ProviderError, RateLimitError, } from "../types/index.js";
 import { logger } from "../utils/logger.js";
+import { createProxyFetch } from "../proxy/proxyFetch.js";
 import { createCohereConfig, getProviderModel, validateApiKey, } from "../utils/providerConfig.js";
-import { composeAbortSignals, createTimeoutController, TimeoutError, } from "../utils/timeout.js";
-import { emitToolEndFromStepFinish } from "../utils/toolEndEmitter.js";
-import { resolveToolChoice } from "../utils/toolChoice.js";
-import { toAnalyticsStreamResult } from "./providerTypeUtils.js";
-import { stepCountIs } from "../utils/tool.js";
-import { streamText } from "../utils/generation.js";
-/**
- * Cohere uses an OpenAI-compatible endpoint at /compatibility/v1 that
- * accepts the same chat-completions shape. Embeddings + Rerank live on
- * the native API and are not exposed through this LLM provider class
- * (use the Cohere SDK directly or the embed/rerank routes when added).
- */
+import { createTimeoutController, TimeoutError } from "../utils/timeout.js";
+import { stripTrailingSlash } from "./openaiChatCompletionsClient.js";
+import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
 const COHERE_DEFAULT_BASE_URL = "https://api.cohere.com/compatibility/v1";
 const getCohereApiKey = () => validateApiKey(createCohereConfig());
 const getDefaultCohereModel = () => getProviderModel("COHERE_MODEL", CohereModels.COMMAND_R_PLUS);
 /**
- * Cohere Provider
+ * Cohere Provider — direct HTTP, no AI SDK.
+ *
+ * Routes Command R / Command R+ chat completions through Cohere's
+ * OpenAI-compatible endpoint at /compatibility/v1. All request/stream/
+ * tool-loop orchestration lives in `OpenAIChatCompletionsProvider`; this
+ * class only declares configuration and provider-specific error mapping.
  *
- * Routes Command R / Command R+ chat completions through Cohere's OpenAI-
- * compatible endpoint. Embed v3 and Rerank v3 are top-tier for RAG but are
- * accessed via the Cohere native SDK / dedicated embedding routes (out of
- * scope for the LLM provider).
+ * Embed v3 is exposed via `embed()` / `embedMany()` backed by a native
+ * POST to /v2/embed (the compatibility path is chat-only).
  *
  * @see https://docs.cohere.com/docs/compatibility-api
+ * @see https://docs.cohere.com/reference/embed
  */
-export class CohereProvider extends BaseProvider {
-    model;
-    apiKey;
-    baseURL;
+export class CohereProvider extends OpenAIChatCompletionsProvider {
     constructor(modelName, sdk, _region, credentials) {
-        const validatedNeurolink = isNeuroLink(sdk) ? sdk : undefined;
-        super(modelName, "cohere", validatedNeurolink);
         const overrideApiKey = credentials?.apiKey?.trim();
-        this.apiKey =
-            overrideApiKey && overrideApiKey.length > 0
-                ? overrideApiKey
-                : getCohereApiKey();
-        this.baseURL =
-            credentials?.baseURL ??
-                process.env.COHERE_BASE_URL ??
-                COHERE_DEFAULT_BASE_URL;
-        const cohere = createOpenAI({
-            apiKey: this.apiKey,
-            baseURL: this.baseURL,
-            fetch: createLoggingFetch("cohere"),
-        });
-        this.model = cohere.chat(this.modelName);
+        const apiKey = overrideApiKey && overrideApiKey.length > 0
+            ? overrideApiKey
+            : getCohereApiKey();
+        const baseURL = credentials?.baseURL?.trim() ||
+            process.env.COHERE_BASE_URL?.trim() ||
+            COHERE_DEFAULT_BASE_URL;
+        super("cohere", modelName, sdk, { baseURL, apiKey });
         logger.debug("Cohere Provider initialized", {
             modelName: this.modelName,
             providerName: this.providerName,
-            baseURL: this.baseURL,
+            baseURL: this.config.baseURL,
         });
     }
-    async executeStream(options, _analysisSchema) {
-        // withClientStreamSpan: keeps the span open until the consumer reaches
-        // end-of-stream / error, so the recorded duration reflects the actual
-        // stream lifetime instead of just setup.
-        return withClientStreamSpan({
-            name: "neurolink.provider.stream",
-            tracer: tracers.provider,
-            attributes: {
-                [ATTR.GEN_AI_SYSTEM]: "cohere",
-                [ATTR.GEN_AI_MODEL]: this.modelName,
-                [ATTR.GEN_AI_OPERATION]: "stream",
-                [ATTR.NL_STREAM_MODE]: true,
-            },
-        }, async () => this.executeStreamInner(options), (r) => r.stream, (r, wrapped) => ({ ...r, stream: wrapped }));
-    }
-    async executeStreamInner(options) {
-        this.validateStreamOptions(options);
-        // Resolve per-call credentials first, then fall back to instance-level.
-        const perCallCreds = options.credentials?.cohere;
-        const effectiveApiKey = perCallCreds?.apiKey?.trim() || this.apiKey;
-        const effectiveBaseURL = perCallCreds?.baseURL || this.baseURL;
-        const startTime = Date.now();
-        const timeout = this.getTimeout(options);
-        const timeoutController = createTimeoutController(timeout, this.providerName, "stream");
-        try {
-            const shouldUseTools = !options.disableTools && this.supportsTools();
-            const tools = shouldUseTools
-                ? options.tools || (await this.getAllTools())
-                : {};
-            const messages = await this.buildMessagesForStream(options);
-            // When per-call credentials differ from instance, build a fresh client.
-            const hasDifferentCreds = effectiveApiKey !== this.apiKey || effectiveBaseURL !== this.baseURL;
-            const model = hasDifferentCreds
-                ? createOpenAI({
-                    apiKey: effectiveApiKey,
-                    baseURL: effectiveBaseURL,
-                    fetch: createLoggingFetch("cohere"),
-                }).chat(this.modelName)
-                : await this.getAISDKModelWithMiddleware(options);
-            const result = await streamText({
-                model,
-                messages,
-                temperature: options.temperature,
-                maxOutputTokens: options.maxTokens,
-                tools,
-                stopWhen: stepCountIs(options.maxSteps || DEFAULT_MAX_STEPS),
-                toolChoice: resolveToolChoice(options, tools, shouldUseTools),
-                abortSignal: composeAbortSignals(options.abortSignal, timeoutController?.controller.signal),
-                experimental_telemetry: this.telemetryHandler.getTelemetryConfig(options),
-                experimental_repairToolCall: this.getToolCallRepairFn(options),
-                onStepFinish: ({ toolCalls, toolResults }) => {
-                    emitToolEndFromStepFinish(this.neurolink?.getEventEmitter(), toolResults);
-                    this.handleToolExecutionStorage(toolCalls, toolResults, options, new Date()).catch((error) => {
-                        logger.warn("[CohereProvider] Failed to store tool executions", {
-                            provider: this.providerName,
-                            error: error instanceof Error ? error.message : String(error),
-                        });
-                    });
-                },
-            });
-            timeoutController?.cleanup();
-            const transformedStream = this.createTextStream(result);
-            const analyticsPromise = streamAnalyticsCollector.createAnalytics(this.providerName, this.modelName, toAnalyticsStreamResult(result), Date.now() - startTime, {
-                requestId: `cohere-stream-${Date.now()}`,
-                streamingMode: true,
-            });
-            return {
-                stream: transformedStream,
-                provider: this.providerName,
-                model: this.modelName,
-                analytics: analyticsPromise,
-                metadata: { startTime, streamId: `cohere-${Date.now()}` },
-            };
-        }
-        catch (error) {
-            timeoutController?.cleanup();
-            throw this.handleProviderError(error);
-        }
-    }
     getProviderName() {
-        return this.providerName;
+        return "cohere";
     }
     getDefaultModel() {
         return getDefaultCohereModel();
     }
-    getAISDKModel() {
-        return this.model;
+    getFallbackModelName() {
+        return CohereModels.COMMAND_R;
+    }
+    getFallbackModels() {
+        return [
+            CohereModels.COMMAND_A,
+            CohereModels.COMMAND_A_REASONING,
+            CohereModels.COMMAND_R_PLUS,
+            CohereModels.COMMAND_R,
+            CohereModels.COMMAND_R7B,
+        ];
     }
     formatProviderError(error) {
         if (error instanceof TimeoutError) {
@@ -176,14 +83,15 @@ export class CohereProvider extends BaseProvider {
         return new ProviderError(`Cohere error: ${message}`, "cohere");
     }
     async validateConfiguration() {
-        return typeof this.apiKey === "string" && this.apiKey.trim().length > 0;
+        return (typeof this.config.apiKey === "string" &&
+            this.config.apiKey.trim().length > 0);
     }
     getConfiguration() {
         return {
             provider: this.providerName,
             model: this.modelName,
             defaultModel: getDefaultCohereModel(),
-            baseURL: this.baseURL,
+            baseURL: this.config.baseURL,
         };
     }
     /**
@@ -210,45 +118,59 @@ export class CohereProvider extends BaseProvider {
     /**
      * Batch embedding via Cohere's native /v2/embed endpoint. Cohere caps at
      * 96 inputs per request; larger batches are chunked.
+     *
+     * Partial failures are not surfaced: if any chunk request fails the whole
+     * call rejects and already-embedded chunks are discarded — callers should
+     * retry the full input.
      */
     async embedMany(texts, modelName) {
         if (texts.length === 0) {
             return [];
         }
         const model = modelName ?? this.getDefaultEmbeddingModel();
-        const baseUrl = this.baseURL.replace(/\/compatibility\/v\d+\/?$/, "");
+        // Strip the compatibility suffix to reach the native API root.
+        const baseUrl = stripTrailingSlash(this.config.baseURL.replace(/\/compatibility\/v\d+\/?$/, ""));
         const url = `${baseUrl}/v2/embed`;
         const BATCH_SIZE = 96;
         const results = [];
+        const fetchImpl = createProxyFetch();
         for (let i = 0; i < texts.length; i += BATCH_SIZE) {
             const batch = texts.slice(i, i + BATCH_SIZE);
-            const response = await fetch(url, {
-                method: "POST",
-                headers: {
-                    Authorization: `Bearer ${this.apiKey}`,
-                    "Content-Type": "application/json",
-                },
-                body: JSON.stringify({
-                    model,
-                    texts: batch,
-                    input_type: "search_document",
-                    embedding_types: ["float"],
-                }),
-            });
-            if (!response.ok) {
-                const body = await response.text();
-                throw this.formatProviderError(new Error(`Cohere /v2/embed failed: ${response.status} — ${body.slice(0, 500)}`));
+            const timeoutController = createTimeoutController(30_000, this.providerName, "generate");
+            try {
+                const response = await fetchImpl(url, {
+                    method: "POST",
+                    headers: {
+                        Authorization: `Bearer ${this.config.apiKey}`,
+                        "Content-Type": "application/json",
+                    },
+                    body: JSON.stringify({
+                        model,
+                        texts: batch,
+                        input_type: "search_document",
+                        embedding_types: ["float"],
+                    }),
+                    ...(timeoutController?.controller.signal
+                        ? { signal: timeoutController.controller.signal }
+                        : {}),
+                });
+                if (!response.ok) {
+                    const body = await response.text().catch(() => "");
+                    throw this.formatProviderError(new Error(`Cohere /v2/embed failed: ${response.status} — ${body.slice(0, 500)}`));
+                }
+                const json = (await response.json());
+                const floatVecs = json.embeddings?.float ??
+                    (Array.isArray(json.embeddings) ? json.embeddings : undefined);
+                if (!floatVecs || floatVecs.length !== batch.length) {
+                    throw new ProviderError(`Cohere /v2/embed returned ${floatVecs?.length ?? 0} embeddings for ${batch.length} inputs.`, "cohere");
+                }
+                results.push(...floatVecs);
             }
-            const json = (await response.json());
-            const floatVecs = json.embeddings?.float ??
-                (Array.isArray(json.embeddings) ? json.embeddings : undefined);
-            if (!floatVecs || floatVecs.length !== batch.length) {
-                throw new ProviderError(`Cohere /v2/embed returned ${floatVecs?.length ?? 0} embeddings for ${batch.length} inputs.`, "cohere");
+            finally {
+                timeoutController?.cleanup();
             }
-            results.push(...floatVecs);
         }
         return results;
     }
 }
-export default CohereProvider;
 //# sourceMappingURL=cohere.js.map

package/dist/lib/providers/groq.d.ts

@@ -1,33 +1,24 @@
 import type { AIProviderName } from "../constants/enums.js";
-import { BaseProvider } from "../core/baseProvider.js";
-import type { NeurolinkCredentials, StreamOptions, StreamResult, ValidationSchema } from "../types/index.js";
-import type { LanguageModel } from "../types/index.js";
+import type { NeurolinkCredentials } from "../types/index.js";
+import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
 /**
- * Groq Provider
+ * Groq Provider — direct HTTP, no AI SDK.
  *
  * Sub-100ms inference of Llama / Mistral / Gemma at api.groq.com/openai/v1
  * (OpenAI-compatible). Best for low-latency tier; trade-off vs other open
  * model hosts is throughput latency, not quality.
  *
+ * All request/stream/tool-loop orchestration lives in
+ * `OpenAIChatCompletionsProvider`; this class only declares configuration
+ * and provider-specific error mapping.
+ *
  * @see https://console.groq.com/docs/quickstart
  */
-export declare class GroqProvider extends BaseProvider {
-    private model;
-    private apiKey;
-    private baseURL;
+export declare class GroqProvider extends OpenAIChatCompletionsProvider {
     constructor(modelName?: string, sdk?: unknown, _region?: string, credentials?: NeurolinkCredentials["groq"]);
-    protected executeStream(options: StreamOptions, _analysisSchema?: ValidationSchema): Promise<StreamResult>;
-    private executeStreamInner;
     protected getProviderName(): AIProviderName;
     protected getDefaultModel(): string;
-    protected getAISDKModel(): LanguageModel;
+    protected getFallbackModelName(): string;
+    protected getFallbackModels(): string[];
     protected formatProviderError(error: unknown): Error;
-    validateConfiguration(): Promise<boolean>;
-    getConfiguration(): {
-        provider: AIProviderName;
-        model: string;
-        defaultModel: string;
-        baseURL: string;
-    };
 }
-export default GroqProvider;

package/dist/lib/providers/groq.js

@@ -1,144 +1,59 @@
-import { createOpenAI } from "@ai-sdk/openai";
 import { GroqModels } from "../constants/enums.js";
-import { BaseProvider } from "../core/baseProvider.js";
-import { DEFAULT_MAX_STEPS } from "../core/constants.js";
-import { streamAnalyticsCollector } from "../core/streamAnalytics.js";
-import { isNeuroLink } from "../neurolink.js";
-import { createLoggingFetch } from "../utils/loggingFetch.js";
-import { tracers, ATTR, withClientStreamSpan } from "../telemetry/index.js";
 import { AuthenticationError, InvalidModelError, ProviderError, RateLimitError, } from "../types/index.js";
 import { logger } from "../utils/logger.js";
 import { createGroqConfig, getProviderModel, validateApiKey, } from "../utils/providerConfig.js";
-import { composeAbortSignals, createTimeoutController, TimeoutError, } from "../utils/timeout.js";
-import { emitToolEndFromStepFinish } from "../utils/toolEndEmitter.js";
-import { resolveToolChoice } from "../utils/toolChoice.js";
-import { toAnalyticsStreamResult } from "./providerTypeUtils.js";
-import { stepCountIs } from "../utils/tool.js";
-import { streamText } from "../utils/generation.js";
+import { TimeoutError } from "../utils/timeout.js";
+import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
 const GROQ_DEFAULT_BASE_URL = "https://api.groq.com/openai/v1";
 const getGroqApiKey = () => validateApiKey(createGroqConfig());
 const getDefaultGroqModel = () => getProviderModel("GROQ_MODEL", GroqModels.LLAMA_3_3_70B_VERSATILE);
 /**
- * Groq Provider
+ * Groq Provider — direct HTTP, no AI SDK.
  *
  * Sub-100ms inference of Llama / Mistral / Gemma at api.groq.com/openai/v1
  * (OpenAI-compatible). Best for low-latency tier; trade-off vs other open
  * model hosts is throughput latency, not quality.
  *
+ * All request/stream/tool-loop orchestration lives in
+ * `OpenAIChatCompletionsProvider`; this class only declares configuration
+ * and provider-specific error mapping.
+ *
  * @see https://console.groq.com/docs/quickstart
  */
-export class GroqProvider extends BaseProvider {
-    model;
-    apiKey;
-    baseURL;
+export class GroqProvider extends OpenAIChatCompletionsProvider {
     constructor(modelName, sdk, _region, credentials) {
-        const validatedNeurolink = isNeuroLink(sdk) ? sdk : undefined;
-        super(modelName, "groq", validatedNeurolink);
         const overrideApiKey = credentials?.apiKey?.trim();
-        this.apiKey =
-            overrideApiKey && overrideApiKey.length > 0
-                ? overrideApiKey
-                : getGroqApiKey();
-        this.baseURL =
-            credentials?.baseURL ??
-                process.env.GROQ_BASE_URL ??
-                GROQ_DEFAULT_BASE_URL;
-        const groq = createOpenAI({
-            apiKey: this.apiKey,
-            baseURL: this.baseURL,
-            fetch: createLoggingFetch("groq"),
-        });
-        this.model = groq.chat(this.modelName);
+        const apiKey = overrideApiKey && overrideApiKey.length > 0
+            ? overrideApiKey
+            : getGroqApiKey();
+        const baseURL = credentials?.baseURL?.trim() ||
+            process.env.GROQ_BASE_URL?.trim() ||
+            GROQ_DEFAULT_BASE_URL;
+        super("groq", modelName, sdk, { baseURL, apiKey });
         logger.debug("Groq Provider initialized", {
             modelName: this.modelName,
             providerName: this.providerName,
-            baseURL: this.baseURL,
+            baseURL: this.config.baseURL,
         });
     }
-    async executeStream(options, _analysisSchema) {
-        return withClientStreamSpan({
-            name: "neurolink.provider.stream",
-            tracer: tracers.provider,
-            attributes: {
-                [ATTR.GEN_AI_SYSTEM]: "groq",
-                [ATTR.GEN_AI_MODEL]: this.modelName,
-                [ATTR.GEN_AI_OPERATION]: "stream",
-                [ATTR.NL_STREAM_MODE]: true,
-            },
-        }, async () => this.executeStreamInner(options), (r) => r.stream, (r, wrapped) => ({ ...r, stream: wrapped }));
-    }
-    async executeStreamInner(options) {
-        this.validateStreamOptions(options);
-        // Resolve per-call credentials first, then fall back to instance-level.
-        const perCallCreds = options.credentials?.groq;
-        const effectiveApiKey = perCallCreds?.apiKey?.trim() || this.apiKey;
-        const effectiveBaseURL = perCallCreds?.baseURL || this.baseURL;
-        const startTime = Date.now();
-        const timeout = this.getTimeout(options);
-        const timeoutController = createTimeoutController(timeout, this.providerName, "stream");
-        try {
-            // Use the canonical BaseProvider helper: merges base tools (MCP/built-in)
-            // with user-provided tools (RAG, etc.) and applies per-call filtering.
-            const shouldUseTools = !options.disableTools && this.supportsTools();
-            const tools = await this.getToolsForStream(options);
-            const messages = await this.buildMessagesForStream(options);
-            // When per-call credentials differ from instance, build a fresh client.
-            const hasDifferentCreds = effectiveApiKey !== this.apiKey || effectiveBaseURL !== this.baseURL;
-            const model = hasDifferentCreds
-                ? createOpenAI({
-                    apiKey: effectiveApiKey,
-                    baseURL: effectiveBaseURL,
-                    fetch: createLoggingFetch("groq"),
-                }).chat(this.modelName)
-                : await this.getAISDKModelWithMiddleware(options);
-            const result = await streamText({
-                model,
-                messages,
-                temperature: options.temperature,
-                maxOutputTokens: options.maxTokens,
-                tools,
-                stopWhen: stepCountIs(options.maxSteps || DEFAULT_MAX_STEPS),
-                toolChoice: resolveToolChoice(options, tools, shouldUseTools),
-                abortSignal: composeAbortSignals(options.abortSignal, timeoutController?.controller.signal),
-                experimental_telemetry: this.telemetryHandler.getTelemetryConfig(options),
-                experimental_repairToolCall: this.getToolCallRepairFn(options),
-                onStepFinish: ({ toolCalls, toolResults }) => {
-                    emitToolEndFromStepFinish(this.neurolink?.getEventEmitter(), toolResults);
-                    this.handleToolExecutionStorage(toolCalls, toolResults, options, new Date()).catch((error) => {
-                        logger.warn("[GroqProvider] Failed to store tool executions", {
-                            provider: this.providerName,
-                            error: error instanceof Error ? error.message : String(error),
-                        });
-                    });
-                },
-            });
-            timeoutController?.cleanup();
-            const transformedStream = this.createTextStream(result);
-            const analyticsPromise = streamAnalyticsCollector.createAnalytics(this.providerName, this.modelName, toAnalyticsStreamResult(result), Date.now() - startTime, {
-                requestId: `groq-stream-${Date.now()}`,
-                streamingMode: true,
-            });
-            return {
-                stream: transformedStream,
-                provider: this.providerName,
-                model: this.modelName,
-                analytics: analyticsPromise,
-                metadata: { startTime, streamId: `groq-${Date.now()}` },
-            };
-        }
-        catch (error) {
-            timeoutController?.cleanup();
-            throw this.handleProviderError(error);
-        }
-    }
     getProviderName() {
-        return this.providerName;
+        return "groq";
     }
     getDefaultModel() {
         return getDefaultGroqModel();
     }
-    getAISDKModel() {
-        return this.model;
+    getFallbackModelName() {
+        return GroqModels.LLAMA_3_1_8B_INSTANT;
+    }
+    getFallbackModels() {
+        return [
+            GroqModels.LLAMA_3_3_70B_VERSATILE,
+            GroqModels.LLAMA_3_1_8B_INSTANT,
+            GroqModels.GEMMA_2_9B_IT,
+            GroqModels.MIXTRAL_8X7B_32768,
+            GroqModels.LLAMA_3_2_90B_VISION_PREVIEW,
+            GroqModels.LLAMA_3_2_11B_VISION_PREVIEW,
+        ];
     }
     formatProviderError(error) {
         if (error instanceof TimeoutError) {
@@ -166,17 +81,5 @@ export class GroqProvider extends BaseProvider {
         }
         return new ProviderError(`Groq error: ${message}`, "groq");
     }
-    async validateConfiguration() {
-        return typeof this.apiKey === "string" && this.apiKey.trim().length > 0;
-    }
-    getConfiguration() {
-        return {
-            provider: this.providerName,
-            model: this.modelName,
-            defaultModel: getDefaultGroqModel(),
-            baseURL: this.baseURL,
-        };
-    }
 }
-export default GroqProvider;
 //# sourceMappingURL=groq.js.map

package/dist/providers/cohere.d.ts

@@ -1,27 +1,26 @@
 import type { AIProviderName } from "../constants/enums.js";
-import { BaseProvider } from "../core/baseProvider.js";
-import type { NeurolinkCredentials, StreamOptions, StreamResult, ValidationSchema } from "../types/index.js";
-import type { LanguageModel } from "../types/index.js";
+import type { NeurolinkCredentials } from "../types/index.js";
+import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
 /**
- * Cohere Provider
+ * Cohere Provider — direct HTTP, no AI SDK.
  *
- * Routes Command R / Command R+ chat completions through Cohere's OpenAI-
- * compatible endpoint. Embed v3 and Rerank v3 are top-tier for RAG but are
- * accessed via the Cohere native SDK / dedicated embedding routes (out of
- * scope for the LLM provider).
+ * Routes Command R / Command R+ chat completions through Cohere's
+ * OpenAI-compatible endpoint at /compatibility/v1. All request/stream/
+ * tool-loop orchestration lives in `OpenAIChatCompletionsProvider`; this
+ * class only declares configuration and provider-specific error mapping.
+ *
+ * Embed v3 is exposed via `embed()` / `embedMany()` backed by a native
+ * POST to /v2/embed (the compatibility path is chat-only).
  *
  * @see https://docs.cohere.com/docs/compatibility-api
+ * @see https://docs.cohere.com/reference/embed
  */
-export declare class CohereProvider extends BaseProvider {
-    private model;
-    private apiKey;
-    private baseURL;
+export declare class CohereProvider extends OpenAIChatCompletionsProvider {
     constructor(modelName?: string, sdk?: unknown, _region?: string, credentials?: NeurolinkCredentials["cohere"]);
-    protected executeStream(options: StreamOptions, _analysisSchema?: ValidationSchema): Promise<StreamResult>;
-    private executeStreamInner;
     protected getProviderName(): AIProviderName;
     protected getDefaultModel(): string;
-    protected getAISDKModel(): LanguageModel;
+    protected getFallbackModelName(): string;
+    protected getFallbackModels(): string[];
     protected formatProviderError(error: unknown): Error;
     validateConfiguration(): Promise<boolean>;
     getConfiguration(): {
@@ -46,7 +45,10 @@ export declare class CohereProvider extends BaseProvider {
     /**
      * Batch embedding via Cohere's native /v2/embed endpoint. Cohere caps at
      * 96 inputs per request; larger batches are chunked.
+     *
+     * Partial failures are not surfaced: if any chunk request fails the whole
+     * call rejects and already-embedded chunks are discarded — callers should
+     * retry the full input.
      */
     embedMany(texts: string[], modelName?: string): Promise<number[][]>;
 }
-export default CohereProvider;