examples/embeddings/index.html

<!DOCTYPE html>
<html lang="en">
<head>
  <meta charset="UTF-8">
  <meta name="viewport" content="width=device-width, initial-scale=1.0">
  <title>wllama.cpp demo</title>

  <style>
    body {
      background-color: rgb(55, 55, 55);
      color: rgb(222, 222, 222);
      font-family: 'Courier New', Courier, monospace;
      padding: 1em;
    }
  </style>
</head>
<body>

  <div id="output"></div>

  <script type="module">
    import { Wllama } from '../../esm/index.js';

    const CONFIG_PATHS = {
      'single-thread/wllama.wasm': '../../esm/single-thread/wllama.wasm',
      'multi-thread/wllama.wasm' : '../../esm/multi-thread/wllama.wasm',
    };
    const MODEL = 'https://huggingface.co/CompendiumLabs/bge-base-en-v1.5-gguf/resolve/main/bge-base-en-v1.5-q4_k_m.gguf';
    const TEXTS = [
      'Taking advantage of the perfect weather, I went out for a leisurely walk through the park.',
      'A stunning sunny day inspired me to get outside and enjoy the beauty of nature while going for a walk.',
      'French is one of her strong points when it comes to languages. Speaking French comes naturally to her.',
      'Represent this sentence for searching relevant passages: Does anyone here speak French?',
    ];

    async function main() {
      let res, tokens, elapsed, buffer;
      const wllama = new Wllama(CONFIG_PATHS);

      print(`DEMO EMBEDDINGS`);
      print(`Loading model ${MODEL}`);
      timeStart();
      await wllama.loadModelFromUrl(MODEL, {
        embeddings: true,
        n_ctx: 1024,
        n_batch: 1024,
        n_ubatch: 1024,
        pooling_type: 'LLAMA_POOLING_TYPE_MEAN',
      });
      print(`Loaded, take ${timeEnd()} ms`);
      print(`BOS token = ${wllama.getBOS()}`);
      print(`EOS token = ${wllama.getEOS()}`);
  
      print(`\n--------------\n`);

      const embeddings = [];
      let i = 0;
      for (const sentence of TEXTS) {
        i++;
        print(`Calculating embedding for sentence #${i}: "${sentence}"`);
        timeStart();
        const vector = await wllama.createEmbedding(sentence);
        print(`OK, take ${timeEnd()} ms`);
        embeddings.push(vector);
      }

      print(`\n--------------\n`);

      const vecDot = (a, b) => a.reduce((acc, _, i) => acc + a[i]*b[i], 0);
      for (let i = 0; i < TEXTS.length - 1; i++) {
        print(`Cosine similarity of sentence #${i+1} and #${i+2} = ${vecDot(embeddings[i], embeddings[i+1])}`);
      }
      
    }

    /////////////////////////////////////////////////////////////////////

    const elemOutput = document.getElementById('output');
    function print(message, bold) {
      const elem = document.createElement('div');
      if (bold) {
        const b = document.createElement('b');
        b.innerText = message;
        elem.appendChild(b);
      } else {
        elem.innerText = message;
      }
      elemOutput.appendChild(elem);
      // scroll to bottom
      setTimeout(() => window.scrollTo({
        top: document.documentElement.scrollHeight - window.innerHeight,
        left: 0,
        behavior: 'smooth',
      }), 10);
    }
    let __startTime = 0;
    function timeStart() {
      __startTime = Date.now();
    }
    function timeEnd() {
      return Date.now() - __startTime;
    }

    main();
  </script>
</body>
</html>