@juspay/neurolink 9.67.2 → 9.67.3

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

@@ -1,15 +1,19 @@
 import type { AIProviderName } from "../constants/enums.js";
-import { BaseProvider } from "../core/baseProvider.js";
-import type { LanguageModel, Schema, StreamOptions, StreamResult, ZodUnknownSchema } from "../types/index.js";
+import type { OpenAICompatBuildBodyArgs, OpenAICompatStreamLifecycleListeners } from "../types/index.js";
+import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
 /**
  * LiteLLM Provider — direct HTTP, no AI SDK. Talks to a LiteLLM proxy
  * server (or any deployment that speaks OpenAI chat-completions + the
  * `/v1/models` and `/v1/embeddings` endpoints).
+ *
+ * All request/stream/tool-loop orchestration lives in
+ * `OpenAIChatCompletionsProvider`. This class adds LiteLLM-specific
+ * behaviour: OTel span wrap with cost (`onStreamStart`), Gemini 2.5
+ * maxTokens skip (`adjustBuildBodyOptions`), ModelAccessDeniedError on
+ * 403, 10-minute model cache (`getAvailableModels`), `LITELLM_FALLBACK_MODELS`
+ * env-driven fallback list, and native `/v1/embeddings`.
  */
-export declare class LiteLLMProvider extends BaseProvider {
-    private config;
-    private credentials?;
-    private resolvedModel?;
+export declare class LiteLLMProvider extends OpenAIChatCompletionsProvider {
     private static modelsCache;
     private static modelsCacheTime;
     private static readonly MODELS_CACHE_DURATION;
@@ -19,30 +23,28 @@ export declare class LiteLLMProvider extends BaseProvider {
     });
     protected getProviderName(): AIProviderName;
     protected getDefaultModel(): string;
+    protected getFallbackModelName(): string;
+    protected getFallbackModels(): string[];
     /**
-     * Abstract from BaseProvider — used by the parent's generate() path which
-     * still goes through `generateText`. Returns a thin LanguageModelV3-shaped
-     * object that delegates to the same HTTP helpers used by executeStream.
+     * Gemini 2.5 models on LiteLLM have a known compatibility issue with
+     * `max_tokens` — strip it before the wire body is built. Applies to
+     * both streaming and non-streaming paths.
      */
-    protected getAISDKModel(): Promise<LanguageModel>;
-    private resolveModelName;
+    protected adjustBuildBodyOptions(modelId: string, opts: OpenAICompatBuildBodyArgs["options"]): OpenAICompatBuildBodyArgs["options"];
     /**
-     * Returns a minimal V3-shaped model. Only used by BaseProvider's
-     * `generate()` non-streaming path which still relies on the parent's
-     * `generateText`. The streaming path bypasses this entirely.
+     * Wrap the stream in an OTel span to capture provider-level latency,
+     * token usage, finish reason, and cost. Matches the pre-migration
+     * behaviour where streamText was wrapped in `neurolink.provider.streamText`.
      */
-    private buildDelegatingModel;
+    protected onStreamStart(modelId: string): OpenAICompatStreamLifecycleListeners | undefined;
     formatProviderError(error: unknown): Error;
-    supportsTools(): boolean;
     /**
-     * Streaming path — drives the LiteLLM proxy directly. No streamText, no
-     * AI SDK orchestrator. Tool calls, multi-step loops, telemetry, abort
-     * handling all inline. OTel span captures gen_ai.system + cost.
+     * Get available models from LiteLLM proxy `/v1/models` endpoint.
+     * Caches results for 10 minutes; falls back to env-driven list or a
+     * minimal safe default if the API fetch fails.
      */
-    protected executeStream(options: StreamOptions, _analysisSchema?: ZodUnknownSchema | Schema<unknown>): Promise<StreamResult>;
-    private runStreamLoop;
-    private streamOneStep;
-    private executeToolBatch;
+    getAvailableModels(): Promise<string[]>;
+    private fetchModelsFromAPI;
     /**
      * Generate an embedding for a single text input via native /v1/embeddings.
      */
@@ -52,13 +54,4 @@ export declare class LiteLLMProvider extends BaseProvider {
      */
     embedMany(texts: string[], modelName?: string): Promise<number[][]>;
     private callEmbeddings;
-    /**
-     * Get available models from LiteLLM proxy `/v1/models` endpoint.
-     * Caches results for 10 minutes; falls back to env-driven list or a
-     * minimal safe default if the API fetch fails.
-     */
-    getAvailableModels(): Promise<string[]>;
-    getFirstAvailableModel(): Promise<string>;
-    private getFallbackModels;
-    private fetchModelsFromAPI;
 }