import { config } from "./config.ts";
import type { EnrichedProduct, LlmVerdict } from "./types.ts";

const SYSTEM_PROMPT = `You are an expert Amazon product analyst specializing in FBA and FBM fulfillment strategy.

Given product data, evaluate each product's viability for selling on Amazon. Consider:

1. **Sales Velocity**: monthlySold and salesRankDrops30 are the most important signals. A product that doesn't sell is worthless regardless of margin. salesRankDrops30 = approximate units sold in 30 days. monthlySold is Keepa's estimate.
2. **Margin Analysis**: Sale price minus unit cost minus fees (FBA or FBM). Aim for >30% ROI minimum. The spreadsheet may include FBA NET and gross profit estimates — cross-check against Keepa pricing data.
3. **Sales Rank (BSR)**: Lower rank = higher demand. Rank <50,000 is good, <1,000 is excellent.
4. **Sales Rank Trend**: Compare current rank vs 90d average. Lower current = improving demand.
5. **Competition**: Number of sellers and Buy Box dynamics. Fewer sellers = easier entry.
6. **Price Stability**: Large price swings (high max vs low min over 90d) = volatile/risky.
7. **FBA vs FBM**: FBA preferred for fast-selling, small/light items. FBM for oversized, slow-moving, or high-margin items where fee savings matter.
8. **MOQ & Capital**: High MOQ with thin margins is risky.
9. **Supply Availability**: Total quantity available from supplier — low stock means limited runway.

Return ONLY a raw JSON array (no markdown, no code fences, no explanation before or after). One verdict per product:
[{ "asin": "B...", "verdict": "FBA" | "FBM" | "SKIP", "confidence": 0-100, "reasoning": "..." }]

Keep each reasoning under 100 characters to stay within output limits.`;

export async function analyzeProducts(
  products: EnrichedProduct[],
): Promise<LlmVerdict[]> {
  const productSummaries = products.map(summarizeForLlm);

  const res = await fetch(`${config.llmUrl}/chat/completions`, {
    method: "POST",
    headers: {
      "Content-Type": "application/json",
      Authorization: "Bearer lm-studio",
    },
    body: JSON.stringify({
      model: config.llmModel,
      messages: [
        { role: "system", content: SYSTEM_PROMPT },
        { role: "user", content: JSON.stringify(productSummaries, null, 2) },
      ],
      temperature: 0.3,
      max_tokens: 2048,
    }),
  });

  if (!res.ok) {
    throw new Error(`LLM API error ${res.status}: ${await res.text()}`);
  }

  const data = (await res.json()) as {
    choices?: { message?: { content?: string } }[];
  };
  const content = data.choices?.[0]?.message?.content ?? "";

  return parseVerdicts(content, products);
}

function summarizeForLlm(p: EnrichedProduct) {
  const salePrice = p.keepa?.currentPrice ?? p.spApi.estimatedSalePrice;
  const referralFee = salePrice * (p.spApi.referralFeePercent / 100);
  const fbaProfit =
    salePrice - p.record.unitCost - p.spApi.fbaFee - referralFee;
  const fbmProfit =
    salePrice - p.record.unitCost - p.spApi.fbmFee - referralFee;

  return {
    asin: p.record.asin,
    name: p.record.name,
    brand: p.record.brand,
    category: p.record.category ?? p.keepa?.categoryTree?.join(" > "),
    unitCost: p.record.unitCost,
    currentPrice: salePrice,
    priceRange90d: p.keepa
      ? {
          min: p.keepa.minPrice90,
          max: p.keepa.maxPrice90,
          avg: p.keepa.avgPrice90,
        }
      : null,
    salesRank: p.keepa?.salesRank ?? p.record.amazonRank,
    salesRankAvg90d: p.keepa?.salesRankAvg90,
    sellerCount: p.keepa?.sellerCount,
    salesVelocity: {
      monthlySold: p.keepa?.monthlySold,
      salesRankDrops30: p.keepa?.salesRankDrops30,
      salesRankDrops90: p.keepa?.salesRankDrops90,
    },
    spreadsheetEstimates: {
      fbaNet: p.record.fbaNet,
      grossProfit: p.record.grossProfit,
      grossProfitPct: p.record.grossProfitPct,
    },
    moq: p.record.moq,
    moqCost: p.record.moqCost,
    totalQtyAvail: p.record.totalQtyAvail,
    fees: {
      fbaFee: p.spApi.fbaFee,
      fbmFee: p.spApi.fbmFee,
      referralFeePercent: p.spApi.referralFeePercent,
      referralFee: Math.round(referralFee * 100) / 100,
    },
    estimatedProfit: {
      fba: Math.round(fbaProfit * 100) / 100,
      fbm: Math.round(fbmProfit * 100) / 100,
    },
    estimatedROI: {
      fba:
        p.record.unitCost > 0
          ? Math.round((fbaProfit / p.record.unitCost) * 100)
          : null,
      fbm:
        p.record.unitCost > 0
          ? Math.round((fbmProfit / p.record.unitCost) * 100)
          : null,
    },
  };
}

function cleanLlmJson(text: string): string {
  // Remove ```json ... ``` or ``` ... ``` wrapping
  const fenceMatch = text.match(/```(?:json)?\s*\n?([\s\S]*?)```/);
  let cleaned = fenceMatch ? fenceMatch[1]!.trim() : text.trim();

  // Fix trailing comma-quote before closing brace: ,"} → "}
  cleaned = cleaned.replace(/,"\s*}/g, '"}');

  // Fix trailing commas before ] or }
  cleaned = cleaned.replace(/,\s*([}\]])/g, "$1");

  return cleaned;
}

function parseVerdicts(
  content: string,
  products: EnrichedProduct[],
): LlmVerdict[] {
  try {
    const cleaned = cleanLlmJson(content);
    const parsed = JSON.parse(cleaned);
    const arr = Array.isArray(parsed)
      ? parsed
      : (parsed.verdicts ?? parsed.results ?? [parsed]);
    return arr.map((v: Record<string, unknown>) => ({
      asin: String(v.asin ?? ""),
      verdict: (["FBA", "FBM", "SKIP"].includes(String(v.verdict))
        ? v.verdict
        : "SKIP") as LlmVerdict["verdict"],
      confidence: typeof v.confidence === "number" ? v.confidence : 0,
      reasoning: String(v.reasoning ?? "No reasoning provided"),
    }));
  } catch (err) {
    console.warn("Failed to parse LLM response, marking all as ANALYSIS_FAILED");
    console.warn("Raw LLM content:", content.slice(0, 500));
    return products.map((p) => ({
      asin: p.record.asin,
      verdict: "SKIP" as const,
      confidence: 0,
      reasoning: `Analysis failed: could not parse LLM output`,
    }));
  }
}