add slop filter

src/router/websocket.ts

@@ -7,9 +7,10 @@ import { validateVideo, validateChannel } from '@/utils/regex'
 import { checkCaptcha, createDatabaseVideo } from '@/utils/common';
 import { downloadVideo } from '@/utils/download';
 import { uploadVideo } from '@/utils/upload';
-import { getChannelVideos } from '@/utils/metadata';
+import { getChannelVideos, getVideo } from '@/utils/metadata';
 import { error } from '@/utils/html'
 import redis from '@/utils/redis';
+import { parseSlop } from '@/utils/slop';
 
 const app = new Elysia()
 const videoIds: Record<string, string> = {}
@@ -128,6 +129,15 @@ app.ws('/save', {
         ws.send('DATA - Captcha validated. Starting download...');
       }
 
+      const data = await getVideo(videoId)
+      const slopScore = await parseSlop(videoId, data.videoDetails.title, 
+        (data.microformat.playerMicroformatRenderer.description?.simpleText || '').replaceAll('\n', '<br>'))
+      
+      if (slopScore >= 4) {
+        sendError(ws, 'Filters can always be wrong. Is the rating wrong? Email me at admin@preservetube.com');
+        return sendError(ws, 'Your download has been rejected by our slop filter.');
+      }
+
       const downloadResult = await downloadVideo(ws, videoId);
       if (downloadResult.fail) {
         await cleanup(ws, videoId);
@@ -201,6 +211,14 @@ app.ws('/savechannel', {
         break;
       }
 
+      const slopScore = await parseSlop(video.video_id, video.title.text, video.description_snippet.text)
+      
+      if (slopScore >= 4) {
+        sendError(ws, 'Filters can always be wrong. Is the rating wrong? Email me at admin@preservetube.com');
+        sendError(ws, 'Your download has been rejected by our slop filter.');
+        continue;
+      }
+
       ws.send(`DATA - Processing video: ${video.title.text}`);
       await redis.set(video.video_id, 'downloading', 'EX', 300);
 

src/utils/slop.ts

@@ -0,0 +1,64 @@
+import redis from '@/utils/redis';
+
+async function analyseSlop(id: string, title: string, description: string) {
+  const llmResponse = await (await fetch('https://nano-gpt.com/api/v1/chat/completions', {
+    method: 'POST',
+    headers: {
+      'Authorization': 'Bearer ' + process.env.NANOGPT_API,
+      'Content-Type': 'application/json'
+    },
+    body: JSON.stringify({
+      model: 'gemini-2.5-flash-lite',
+      messages: [
+        { role: 'user', content: `Role: You are "The Slop Detector." Analyze video titles and rate them 0-5 on the "Slop Score." Slop is derivative content using popular movie/TV footage + trendy audio + basic editing.
+
+### SLOP SCORE (0-5)
+- **0:** Original, complex, artistic. No slop signals.
+- **1:** Fan edit with unique perspective, non-obvious choices.
+- **2:** Well-made but predictable, adds nothing new.
+- **3:** Some slop signals present, minimal effort visible.
+- **4:** Textbook slop — popular character + overused/slowed song + 4K tag + "Edit."
+- **5:** Maximum slop density — every possible signal stacked.
+
+### SLOP SIGNALS
+Use these as intuition guides, not a checklist. Weight them by how many stack together.
+
+- **Emoji in title** — strong signal, especially 😂🤣😎😍🥺. Multiple emoji = very strong.
+- **Title is entirely hashtags** — near-instant slop.
+- **Unicode styled text** (𝐋𝐢𝐤𝐞 𝐓𝐡𝐢𝐬) — common in character edits.
+- **Pipe separators** (|) splitting title into: caption | source | song — classic slop structure.
+- **4K / [4K] tag** — almost always slop when paired with anything else.
+- **"Edit" / "OneShot" / "Morphosis"** suffix or delimiter usage (║ Edit ║, 「Edit」).
+- **Slowed / Reverb / Slowed+Reverb / MONTAGEM** — audio slop markers.
+- **Known slop franchises:** Johnny English, Mr. Bean, Breaking Bad, Peaky Blinders, American Psycho, Patrick Bateman, The Boys, Homelander, Dexter, Joe Goldberg, Squid Game, Rick Grimes, Thomas Shelby, John Wick, Kingsman, and similar.
+- **Song name explicitly in title** — especially if slowed/remixed.
+- **Description** — if it contains hashtag spam, "subscribe," "no copyright," or music credits it reinforces slop signals from the title.
+- **Tutorials and how-to videos** — score 0, always.
+
+### OUTPUT
+Valid JSON only. No other text. Reasoning max one sentence, and brief.
+
+{"score": 0, "reasoning": "..."}
+
+User Title: ${title}
+User Description: ${description.slice(0,100)}` }
+      ]
+    })
+  })).json() 
+
+  const parsedResponse: {score: number, reasoning: string} = JSON.parse(llmResponse.choices[0].message.content.replace(/```json|```/g, '').trim())
+  console.log(`parsed ${id} - ${JSON.stringify(parsedResponse)}`)
+
+  return parsedResponse
+}
+
+async function parseSlop(id: string, title: string, description: string): Promise<number> {
+  const cachedSlop = await redis.get(`slop:${id}`)
+  if (cachedSlop) return parseInt(cachedSlop)
+
+  const { score } = await analyseSlop(id , title, description)
+  await redis.set(`slop:${id}`, score, 'EX', 60 * 60 * 24 * 7)
+  return score
+}
+
+export { parseSlop }