import type { LlmClient } from '@/lib/ai/llm-client';
import { ChatExtractionSchema } from '@/lib/ai/chat-extraction-types';
import type { ChatExtractionData } from '@/lib/ai/chat-extraction-types';
import type { KnowledgeItem } from '@/lib/types/knowledge';

const SYSTEM_PROMPT = `
You are the Product Chat Signal Extractor for stalled SaaS projects.
- Read the provided transcript carefully.
- Extract grounded signals about the product, market, users, execution status, and unknowns.
- Never invent data. Use "null" or empty arrays when the transcript lacks information.
- Respond with valid JSON that matches the provided schema exactly. Do not include prose or code fences.
`.trim();

export async function runChatExtraction(
  knowledgeItem: KnowledgeItem,
  llm: LlmClient,
): Promise<ChatExtractionData> {
  const transcript = knowledgeItem.content.trim();

  const userMessage = `
You will analyze the following transcript. Use message references when listing evidence (e.g., msg_1). 
Focus on actionable product-building insights.

TRANSCRIPT_START
${transcript}
TRANSCRIPT_END`.trim();

  return llm.structuredCall<ChatExtractionData>({
    model: 'gemini',
    systemPrompt: SYSTEM_PROMPT,
    messages: [
      {
        role: 'user',
        content: userMessage,
      },
    ],
    schema: ChatExtractionSchema,
    temperature: 0.2,
  });
}