Momento/memento-note/lib/ai/providers/deepseek.ts

import { createOpenAI } from '@ai-sdk/openai';
import { generateObject, generateText as aiGenerateText, embed, stepCountIs } from 'ai';
import { z } from 'zod';
import { AIProvider, TagSuggestion, TitleSuggestion, ToolUseOptions, ToolCallResult } from '../types';

export class DeepSeekProvider implements AIProvider {
  private model: any;
  private embeddingModel: any;

  constructor(apiKey: string, modelName: string = 'deepseek-chat', embeddingModelName: string = 'deepseek-embedding') {
    // Create OpenAI-compatible client for DeepSeek
    // Disable extended thinking to ensure reliable tool/function calling
    const deepseek = createOpenAI({
      baseURL: 'https://api.deepseek.com/v1',
      apiKey: apiKey,
      fetch: async (url, options) => {
        if (options?.body) {
          try {
            const body = JSON.parse(options.body as string)
            // Disable thinking mode — tool calling is unreliable with it enabled
            body.thinking = { type: 'disabled' }
            return fetch(url, { ...options, body: JSON.stringify(body) })
          } catch { /* ignore parse errors */ }
        }
        return fetch(url, options)
      },
    });

    this.model = deepseek.chat(modelName);
    this.embeddingModel = deepseek.embedding(embeddingModelName);
  }

  async generateTags(content: string, language?: string): Promise<TagSuggestion[]> {
    try {
      // DeepSeek doesn't support response_format: json_schema — use generateText + manual parse
      const { text } = await aiGenerateText({
        model: this.model,
        prompt: `Analyze the following note and suggest 1 to 5 relevant tags as a JSON array.
Return ONLY a JSON array like: [{"tag":"example","confidence":0.9}]
Note content: "${content.substring(0, 1500)}"`,
      });

      const clean = text.replace(/^```json\n?/, '').replace(/\n?```$/, '').trim();
      const parsed = JSON.parse(clean);
      const arr = Array.isArray(parsed) ? parsed : (parsed.tags || []);
      return arr.map((t: any) => ({
        tag: t.tag || t.label || t.name || '',
        confidence: t.confidence || t.score || 0.7,
      }));
    } catch (e) {
      console.error('Error generating tags (DeepSeek):', e);
      return [];
    }
  }

  async getEmbeddings(text: string): Promise<number[]> {
    try {
      const { embedding } = await embed({
        model: this.embeddingModel,
        value: text,
      });
      return embedding;
    } catch (e) {
      console.error('Error generating embeddings (DeepSeek):', e);
      throw e;
    }
  }

  async generateTitles(prompt: string): Promise<TitleSuggestion[]> {
    try {
      const { object } = await generateObject({
        model: this.model,
        schema: z.object({
          titles: z.array(z.object({
            title: z.string().describe('Suggested title'),
            confidence: z.number().min(0).max(1).describe('Confidence level between 0 and 1')
          }))
        }),
        prompt: prompt,
      });

      return object.titles;
    } catch (e) {
      console.error('Error generating titles (DeepSeek):', e);
      return [];
    }
  }

  async generateText(prompt: string): Promise<string> {
    try {
      const { text } = await aiGenerateText({
        model: this.model,
        prompt: prompt,
      });

      return text.trim();
    } catch (e) {
      console.error('Error generating text (DeepSeek):', e);
      throw e;
    }
  }

  async chat(messages: any[], systemPrompt?: string): Promise<any> {
    try {
      const { text } = await aiGenerateText({
        model: this.model,
        system: systemPrompt,
        messages: messages,
      });

      return { text: text.trim() };
    } catch (e) {
      console.error('Error in chat (DeepSeek):', e);
      throw e;
    }
  }

  async generateWithTools(options: ToolUseOptions): Promise<ToolCallResult> {
    const { tools, maxSteps = 10, systemPrompt, messages, prompt } = options

    const buildOpts = (steps: number): Record<string, any> => {
      const opts: Record<string, any> = { model: this.model, tools, stopWhen: stepCountIs(steps) }
      if (systemPrompt) opts.system = systemPrompt
      if (messages) opts.messages = messages
      else if (prompt) opts.prompt = prompt
      return opts
    }

    const toResult = (r: any): ToolCallResult => ({
      toolCalls: r.toolCalls?.map((tc: any) => ({ toolName: tc.toolName, input: tc.input })) || [],
      toolResults: r.toolResults?.map((tr: any) => ({ toolName: tr.toolName, input: tr.input, output: tr.output })) || [],
      text: r.text,
      steps: r.steps?.map((step: any) => ({
        text: step.text,
        toolCalls: step.toolCalls?.map((tc: any) => ({ toolName: tc.toolName, input: tc.input })) || [],
        toolResults: step.toolResults?.map((tr: any) => ({ toolName: tr.toolName, input: tr.input, output: tr.output })) || [],
      })) || [],
    })

    try {
      const result = await aiGenerateText(buildOpts(maxSteps) as any)
      return toResult(result)
    } catch (err: any) {
      // DeepSeek reasoning/thinking models require reasoning_content to be passed back
      // between multi-step calls, which the AI SDK doesn't handle automatically.
      // Retry with a single step so the model calls the tool directly without multi-turn.
      const msg: string = err?.message || String(err)
      if (msg.includes('reasoning_content') || msg.includes('thinking mode')) {
        console.warn('[DeepSeek] Reasoning model detected — retrying with maxSteps=1')
        const result = await aiGenerateText(buildOpts(1) as any)
        return toResult(result)
      }
      throw err
    }
  }

  getModel() {
    return this.model;
  }
}