import { OpenAIClient } from '@langchain/openai';
import { AIModelEntity, ICopilotModel } from '@metad/contracts';
import { ChatOAICompatReasoningModel, LargeLanguageModel, TChatModelOptions } from '@xpert-ai/plugin-sdk';
import { AIMessage, AIMessageChunk, BaseMessage } from '@langchain/core/messages';
import { ChatGenerationChunk } from '@langchain/core/outputs';
import { VLLMProviderStrategy } from '../provider.strategy.js';
import { VLLMModelCredentials } from '../types.js';
/**
 * vLLM-specific chat model that extracts reasoning content from redacted_reasoning tags
 * and sets it in additional_kwargs.reasoning_content for proper display
 * Key fix: Override _convertCompletionsDeltaToBaseMessageChunk to handle tags in delta content
 * This ensures reasoning chunks have empty content, so platform code can recognize reasoning type
 */
declare class VLLMChatOAICompatReasoningModel extends ChatOAICompatReasoningModel {
    private accumulatedReasoning;
    private finalReasoningContent;
    private inReasoningMode;
    private reasoningComplete;
    private readonly thinkingEnabled;
    constructor(fields: any);
    /**
     * Override _convertCompletionsDeltaToBaseMessageChunk to handle vLLM's tag format
     * This is the key fix: process tags at delta conversion stage, not in streaming
     * vLLM format: "推理内容</think>正常回复" (tag mixed with content)
     * Platform logic: if content exists, return type: "text" (only checks reasoning_content if content is empty)
     * Solution: set content to empty for reasoning chunks, populate reasoning_content
     */
    protected _convertCompletionsDeltaToBaseMessageChunk(delta: Record<string, any>, rawResponse: OpenAIClient.ChatCompletionChunk, defaultRole?: 'function' | 'user' | 'system' | 'developer' | 'assistant' | 'tool'): AIMessageChunk;
    /**
     * Override _convertCompletionsMessageToBaseMessage for non-streaming case
     * Handle reasoning content from tags in final message
     */
    protected _convertCompletionsMessageToBaseMessage(message: OpenAIClient.ChatCompletionMessage, rawResponse: OpenAIClient.ChatCompletion): AIMessage;
    /**
     * Override _generate to extract reasoning content from tags after generation
     * This handles non-streaming case
     */
    _generate(messages: BaseMessage[], options?: Parameters<ChatOAICompatReasoningModel['_generate']>[1], runManager?: Parameters<ChatOAICompatReasoningModel['_generate']>[2]): Promise<import("@langchain/core/outputs").ChatResult>;
    /**
     * Override streaming to reset state for each new stream
     * The actual tag processing is done in _convertCompletionsDeltaToBaseMessageChunk
     */
    _streamResponseChunks(messages: BaseMessage[], options?: Parameters<ChatOAICompatReasoningModel['_streamResponseChunks']>[1], runManager?: Parameters<ChatOAICompatReasoningModel['_streamResponseChunks']>[2]): AsyncGenerator<ChatGenerationChunk, void, unknown>;
}
export declare class VLLMLargeLanguageModel extends LargeLanguageModel {
    #private;
    constructor(modelProvider: VLLMProviderStrategy);
    validateCredentials(model: string, credentials: VLLMModelCredentials): Promise<void>;
    getChatModel(copilotModel: ICopilotModel, options?: TChatModelOptions): VLLMChatOAICompatReasoningModel;
    /**
     * Generate model schema from credentials for customizable models
     * This method dynamically generates parameter rules including thinking mode
     * Merges parent class parameter rules (streaming, temperature, etc.) with thinking mode
     */
    getCustomizableModelSchemaFromCredentials(model: string, credentials: Record<string, any>): AIModelEntity | null;
}
export {};
//# sourceMappingURL=llm.d.ts.map