Sentiment analysis

// recipe-sentiment.ts
// Harvests comments about a topic, builds an LLM-ready classification batch.
// Run with: SOCIALCRAWL_KEY=sc_... npx tsx recipe-sentiment.ts

const KEY = process.env.SOCIALCRAWL_KEY;
if (!KEY) throw new Error("Set SOCIALCRAWL_KEY");

const BASE = "https://www.socialcrawl.dev/v1";
const topic = "github copilot pricing";

async function get(path: string, params: Record<string, string>) {
  const url = new URL(`${BASE}/${path}`);
  for (const [k, v] of Object.entries(params)) url.searchParams.set(k, v);
  const res = await fetch(url, {
    headers: { "x-api-key": KEY!, accept: "application/json" },
  });
  return (await res.json()) as {
    success: boolean;
    data?: Record<string, any>;
  };
}

// ── Step 1: Find the conversation (comments come pre-attached) ────────────
const search = await get("search/everywhere", {
  query: topic,
  lookback_days: "14",
});
if (!search.success) throw new Error("search failed");

type Sample = { platform: string; post_url: string; text: string };
const samples: Sample[] = [];

for (const item of search.data!.items as Array<{
  source: string;
  url: string;
  source_items: Array<{
    metadata?: { top_comments?: Array<{ text: string }> };
  }>;
}>) {
  for (const c of item.source_items[0]?.metadata?.top_comments ?? []) {
    samples.push({ platform: item.source, post_url: item.url, text: c.text });
  }
}

// ── Step 2: Deepen the sample on the 3 hottest posts ──────────────────────
const COMMENTS_PATH: Record<string, string> = {
  tiktok: "tiktok/post/comments",
  youtube: "youtube/video/comments",
  reddit: "reddit/post/comments",
};

const hot = (search.data!.items as Array<{ source: string; url: string }>)
  .filter((i) => COMMENTS_PATH[i.source])
  .slice(0, 3);

for (const post of hot) {
  const res = await get(COMMENTS_PATH[post.source]!, { url: post.url });
  for (const item of (res.data?.items ?? []) as Array<{
    comment?: { text?: string | null };
  }>) {
    if (item.comment?.text)
      samples.push({
        platform: post.source,
        post_url: post.url,
        text: item.comment.text,
      });
  }
}

console.log(`${samples.length} comments harvested about "${topic}"`);

// ── Step 3: Hand the batch to your LLM of choice ──────────────────────────
const prompt = `Classify the sentiment of each social media comment about
"${topic}" as "positive", "negative", or "neutral". Judge sentiment toward
the topic itself, not general mood. Account for sarcasm. Return a JSON array
of { index, sentiment, confidence } objects.

Comments:
${samples.map((s, i) => `${i}. [${s.platform}] ${s.text.slice(0, 280)}`).join("\n")}`;

// Send `prompt` to any LLM, OpenAI, Anthropic, Gemini, a local model.
// Then aggregate: share positive vs negative per platform, over time, etc.
console.log(prompt.slice(0, 600), "…");

// `samples` after Step 2, the LLM input batch:
[
  { "platform": "reddit", "post_url": "https://reddit.com/r/programming/...", "text": "Honestly worth every cent, it writes half my tests." },
  { "platform": "youtube", "post_url": "https://www.youtube.com/watch?v=...", "text": "$39/mo for autocomplete, lol no thanks" },
  { "platform": "tiktok", "post_url": "https://www.tiktok.com/@.../video/...", "text": "the new pricing killed it for students fr" },
  // ... 80-200 more, depending on how active the topic is
]

// And the LLM's output after Step 3:
[
  { "index": 0, "sentiment": "positive", "confidence": 0.95 },
  { "index": 1, "sentiment": "negative", "confidence": 0.9 },
  { "index": 2, "sentiment": "negative", "confidence": 0.85 },
]