llmgw/internal/web/templates/chat.html

{{define "chat"}}<!doctype html>
<html lang="en"><head>
<meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
<title>Chat · llmgw</title><link rel="stylesheet" href="/_static/base.css">
<link rel="stylesheet" href="/_static/katex/katex.min.css">
<script defer src="/_static/katex/katex.min.js"></script>
<script defer src="/_static/katex/auto-render.min.js"></script>
<meta name="csrf" content="{{.CSRF}}">
</head><body><div class="wrap">
<header>
  <h1>Chat</h1>
  <p class="sub">{{.User.Email}} · project {{.User.ProjectID}}</p>
  <nav>
    <a href="/profile">your keys</a>
    {{if .User.IsAdmin}}<a href="/admin">cluster</a>{{end}}
    <a href="/">about this gateway</a>
  </nav>
</header>

{{if .Models}}
<div class="chat-bar">
  <label class="inline-label">Model
    <select id="model">{{range .Models}}<option value="{{.Name}}" data-simulated="{{if .Simulated}}1{{else}}0{{end}}">{{.Name}}{{if .Simulated}} (simulated){{end}}</option>{{end}}</select>
  </label>
  <button id="reset" type="button" class="link">New conversation</button>
</div>

<p id="sim-note" class="note" hidden>
  This model is simulated: it exercises scheduling and reports timings, and its
  replies are placeholder text, not inference.
</p>

<div id="log" class="chat-log" aria-live="polite"></div>

<form id="composer" class="chat-composer">
  <textarea id="input" rows="3" placeholder="Ask something…" required></textarea>
  <button type="submit" id="send">Send</button>
</form>
<p class="note">Nothing here is stored. The conversation lives in this page and
is gone when you close it — what is recorded is the same usage any request
records: tokens, cost and latency.</p>
{{else}}
<p class="empty">No models are available to you right now.</p>
<p class="note">Either nothing is placed yet, or your project's allowlist does
not include anything that is. <code>/admin</code> shows the cluster if you have
access to it.</p>
{{end}}

<footer><span>llmgw {{.Version}}</span><span>{{.Now}}</span></footer>
</div>

<script>
// Vanilla, inline, no build step — the same reason the charts are hand-drawn
// SVG. It streams because a chat that hangs for ten seconds and then dumps a
// paragraph feels broken, and on CPU inference ten seconds is normal.
(function () {
  const log = document.getElementById("log");
  if (!log) return;
  const form = document.getElementById("composer");
  const input = document.getElementById("input");
  const send = document.getElementById("send");
  const model = document.getElementById("model");
  // From a meta tag rather than interpolated into the script, so the token
  // travels through the same HTML escaping as everything else on the page.
  const csrf = document.querySelector('meta[name="csrf"]').content;
  let messages = [];

  document.getElementById("reset").onclick = () => { messages = []; log.textContent = ""; };

  // The note follows the picker rather than the reply: someone should know what
  // they are about to get before they spend a minute waiting for it.
  const simNote = document.getElementById("sim-note");
  function showSimNote() {
    const opt = model.selectedOptions[0];
    simNote.hidden = !opt || opt.dataset.simulated !== "1";
  }
  model.addEventListener("change", showSimNote);
  showSimNote();

  function bubble(role) {
    const el = document.createElement("div");
    el.className = "bubble " + role;
    // textContent throughout: the reply is model output and the prompt is user
    // input, and neither has any business being parsed as markup.
    log.appendChild(el);
    log.scrollTop = log.scrollHeight;
    return el;
  }

  // Maths is rendered after the text is already in place as text.
  //
  // That order is the whole safety argument. renderMathInElement walks the text
  // nodes of an element whose content was set with textContent, so nothing the
  // model wrote is ever parsed as HTML — only the spans KaTeX recognises become
  // markup, and that markup is KaTeX's, not the model's.
  //
  // trust stays false, which is KaTeX's default and worth stating: it is what
  // keeps \href, \url and \includegraphics out of a reply.
  function renderMath(el) {
    if (typeof renderMathInElement !== "function") return;
    try {
      renderMathInElement(el, {
        delimiters: [
          { left: "$$", right: "$$", display: true },
          { left: "\\[", right: "\\]", display: true },
          { left: "\\(", right: "\\)", display: false },
          // Single $ last: a model writing about prices produces "$5 and $10",
          // and treating that as maths turns a sentence into gibberish. It is
          // still enabled because models overwhelmingly use it for maths, and
          // ignoreOnError leaves anything that does not parse as plain text.
          { left: "$", right: "$", display: false },
        ],
        throwOnError: false,
        errorColor: "var(--bad)",
        trust: false,
      });
    } catch (e) { /* a reply without maths is better than no reply */ }
  }

  form.onsubmit = async (e) => {
    e.preventDefault();
    const text = input.value.trim();
    if (!text) return;

    input.value = "";
    send.disabled = true;
    bubble("you").textContent = text;
    messages.push({ role: "user", content: text });

    const out = bubble("assistant");
    let reply = "";
    try {
      const resp = await fetch("/chat/send", {
        method: "POST",
        headers: { "Content-Type": "application/json", "X-CSRF-Token": csrf },
        body: JSON.stringify({ model: model.value, messages: messages }),
      });
      if (!resp.ok) {
        out.textContent = "(" + resp.status + ") " + (await resp.text()).slice(0, 400);
        out.classList.add("failed");
        send.disabled = false;
        return;
      }

      const reader = resp.body.getReader();
      const dec = new TextDecoder();
      let buf = "";
      for (;;) {
        const { done, value } = await reader.read();
        if (done) break;
        buf += dec.decode(value, { stream: true });
        // Server-sent events arrive in whatever chunks the network felt like;
        // only whole lines can be parsed.
        let nl;
        while ((nl = buf.indexOf("\n")) >= 0) {
          const line = buf.slice(0, nl).trim();
          buf = buf.slice(nl + 1);
          if (!line.startsWith("data:")) continue;
          const payload = line.slice(5).trim();
          if (payload === "[DONE]") continue;
          try {
            const chunk = JSON.parse(payload);
            const piece = chunk.choices?.[0]?.delta?.content;
            if (piece) { reply += piece; out.textContent = reply; log.scrollTop = log.scrollHeight; }
          } catch (_) { /* a partial frame; the next read completes it */ }
        }
      }
      if (reply) {
        messages.push({ role: "assistant", content: reply });
        // Once, at the end. Re-rendering every chunk would try to typeset half
        // an expression on each token and flicker its way through the reply.
        renderMath(out);
      }
      else { out.textContent = "(no content came back)"; out.classList.add("failed"); }
    } catch (err) {
      out.textContent = "(" + err + ")";
      out.classList.add("failed");
    } finally {
      send.disabled = false;
      input.focus();
    }
  };
})();
</script>
</body></html>{{end}}