Momento/memento-note/lib/ai/providers/custom-openai.ts

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

export class CustomOpenAIProvider implements AIProvider {
  private model: any;
  private apiKey: string;
  private baseUrl: string;
  private embeddingModelName: string;

  constructor(
    apiKey: string,
    baseUrl: string,
    modelName: string = 'gpt-4o-mini',
    embeddingModelName: string = 'text-embedding-3-small'
  ) {
    this.apiKey = apiKey;
    this.baseUrl = baseUrl.endsWith('/') ? baseUrl.slice(0, -1) : baseUrl;
    this.embeddingModelName = embeddingModelName;
    // Create OpenAI-compatible client with custom base URL
    // Use .chat() to force /chat/completions endpoint (avoids Responses API)
    const customClient = createOpenAI({
      baseURL: baseUrl,
      apiKey: apiKey,
      fetch: async (url, options) => {
        const headers = new Headers(options?.headers);
        headers.set('HTTP-Referer', 'https://localhost:3000');
        headers.set('X-Title', 'Memento AI');
        // Disable DeepSeek extended thinking for reliable tool/function calling
        if (options?.body) {
          try {
            const body = JSON.parse(options.body as string)
            if (
              typeof body.model === 'string' &&
              (body.model.includes('deepseek') || body.model.includes('thinking') || body.model.includes('reasoner'))
            ) {
              body.thinking = { type: 'disabled' }
            }
            return fetch(url, { ...options, headers, body: JSON.stringify(body) })
          } catch { /* ignore parse errors */ }
        }
        return fetch(url, { ...options, headers });
      }
    });

    this.model = customClient.chat(modelName);
  }

  private async fetchWithTimeout(url: string, options: RequestInit, timeoutMs: number = 60_000): Promise<Response> {
    const controller = new AbortController()
    const timer = setTimeout(() => controller.abort(), timeoutMs)
    try {
      return await fetch(url, { ...options, signal: controller.signal })
    } finally {
      clearTimeout(timer)
    }
  }

  async generateTags(content: string): Promise<TagSuggestion[]> {
    try {
      const { object } = await generateObject({
        model: this.model,
        schema: z.object({
          tags: z.array(z.object({
            tag: z.string().describe('Short tag name in lowercase'),
            confidence: z.number().min(0).max(1).describe('Confidence level between 0 and 1')
          }))
        }),
        prompt: `Analyze the following note and suggest 1 to 5 relevant tags.
        Note content: "${content}"`,
      });

      return object.tags;
    } catch (e) {
      console.error('Error generating tags (Custom OpenAI):', e);
      return [];
    }
  }

  async getEmbeddings(text: string): Promise<number[]> {
    try {
      const response = await this.fetchWithTimeout(`${this.baseUrl}/embeddings`, {
        method: 'POST',
        headers: {
          'Content-Type': 'application/json',
          'Authorization': `Bearer ${this.apiKey}`,
          'HTTP-Referer': 'https://localhost:3000',
          'X-Title': 'Memento AI',
        },
        body: JSON.stringify({
          model: this.embeddingModelName,
          input: text,
        }),
      });

      if (!response.ok) {
        const errText = await response.text();
        throw new Error(`${this.baseUrl}/embeddings error ${response.status}: ${errText}`);
      }

      const data = await response.json();

      // Standard OpenAI-compatible response: { data: [{ embedding: number[] }] }
      if (data.data && Array.isArray(data.data) && data.data[0]?.embedding) {
        return data.data[0].embedding;
      }

      // Fallback: some providers return { embedding: number[] }
      if (data.embedding && Array.isArray(data.embedding)) {
        return data.embedding;
      }

      throw new Error(`Unexpected embeddings response shape: ${JSON.stringify(data)}`);
    } catch (e) {
      console.error('Error generating embeddings (CustomOpenAI):', e);
      throw e;
    }
  }

  async generateTitles(prompt: string): Promise<TitleSuggestion[]> {
    try {
      // Use generateText instead of generateObject — DeepSeek doesn't support
      // response_format: json_schema via the OpenAI compat layer
      const { text } = await aiGenerateText({
        model: this.model,
        prompt: prompt,
      })

      // Parse the JSON array from the text response — strip markdown code fences if present
      const parsed = JSON.parse(text.replace(/^```json\n?/,'').replace(/\n?```$/,'').trim())
      const titles = Array.isArray(parsed) ? parsed : (parsed.titles || parsed.suggestions || [])
      return titles.map((t: any) => ({
        title: typeof t === 'string' ? t : t.title || t.name || '',
        confidence: typeof t === 'number' ? t : (t.confidence || t.score || 0.5),
      }))
    } catch (e) {
      console.error('Error generating titles (Custom OpenAI):', 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 (Custom OpenAI):', 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 (Custom OpenAI):', 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 via the OpenAI-compat layer.
      // Retry with a single step so the model calls the tool directly.
      const msg: string = err?.message || String(err)
      if (msg.includes('reasoning_content') || msg.includes('thinking mode')) {
        console.warn('[CustomOpenAI] Reasoning model detected — retrying with maxSteps=1')
        const result = await aiGenerateText(buildOpts(1) as any)
        return toResult(result)
      }
      throw err
    }
  }

  getModel() {
    return this.model;
  }
}