Caption Summaries #

// For this demo, `videos` is a list of VTT file URLs, grab value by index
const url = videos[parseInt(id)] ?? null;

// Load that VTT file from Stream:
const captions = await fetch(url).then(res => res.text(), res => `Could not fetch captions: ${res.status} ${res.statusText}`);

// Pull and concatenate the text portion of all the cues
const [meta, ...stack] = captions.split('\n\n');
const plaintext = stack.reduce((acc, cue) => {
  const [id, time, ...text] = cue.split('\n');
  return acc.concat(' ', text.join(' '));
}, '');

let summary;

// Send the text to the model and return the response.
switch (model) {
  case "bart":
    // Not really a summary, more like a truncated slight shortening
    summary = await env.AI.run('@cf/facebook/bart-large-cnn', {
      input_text: plaintext,
      max_length: 512,
    });
    break;
  case "llama":
    // Much more through / conversational summary
    const messages = [
      {role: 'system', content: 'Summarize this video transcript'},
      {role: 'user', content: plaintext},
    ];

    summary = await env.AI.run("@cf/meta/llama-3.3-70b-instruct-fp8-fast",
      {
        messages,
        stream: false,
        max_tokens: 256,
      }
    );
    break;
}

return new Response(JSON.stringify(summary, null, 2));