Search

  - `message`, `explanation_short`, and `evidence[]`
  The prompt prefers identity‑first summaries and concrete evidence (role/title, org, city) and penalizes vague name‑only matches.
- Browser Search (model: `openai/gpt-oss-20b` + `browser_search` tool): performs targeted queries (quoted email, handle+domain, site:linkedin, site:github, org/staff pages). Produces an analysis that is converted to JSON with status/message/explanation/evidence.
3) Final judge (model: `openai/gpt-oss-120b`)
Consolidates assessor outputs + heuristics (academic, role, domain). Policy:
- Severity order: `spam > person_none > person_low > person_high`.

    const curl = [
      "curl --request POST \\\n",
      "    --url https://api.groq.com/openai/v1/chat/completions \\\n",
      "    --header 'authorization: Bearer " + apiKey + "' \\\n",
      "    --header 'content-type: application/json' \\\n",

export async function finalJudgeAssess(apiKey, email, systemPrompt, evidence) {
  console.log('>> [final judge] evidence:', email, JSON.stringify(evidence));
  const model = 'openai/gpt-oss-120b';
  const schema = '{ "status": "person_high|person_low|person_none|spam", "message": string, "explanation_short": string, "evidence": string[] }';
  const sys = [

async function convertAnalysisToJson(apiKey, email, analysisText) {
  try {
    const convModel = 'openai/gpt-oss-20b';
    const convSys = [
      'You convert analysis text into a strict JSON object. Do not browse or call tools.',
    { role: 'user', content: 'Email: ' + email + '\nTask: web search legitimacy check.' }
  ];
  const model = 'openai/gpt-oss-20b';
  const payload = { model, messages, stream: false, temperature: 0.2, tools: [{ type: 'browser_search' }] };
  const r1 = await groqChatCompletion(apiKey, payload);

const VOICE = "cedar";
const OPENAI_API_KEY = Deno.env.get("OPENAI_API_KEY");
if (!OPENAI_API_KEY) {
  throw new Error("🔴 OpenAI API key not configured");
}
export function makeHeaders(contentType?: string) {
  const obj: Record<string, string> = {
    Authorization: `Bearer ${OPENAI_API_KEY}`,
  };
  if (contentType) obj["Content-Type"] = contentType;

sip.post("/", async (c) => {
  // Verify the webhook.
  const OPENAI_SIGNING_SECRET = Deno.env.get("OPENAI_SIGNING_SECRET");
  if (!OPENAI_SIGNING_SECRET) {
    console.error("🔴 webhook secret not configured");
    return c.text("Internal error", 500);
  }
  const webhook = new Webhook(OPENAI_SIGNING_SECRET);
  const bodyStr = await c.req.text();
  let callId: string | undefined;
  // Accept the call.
  const url = `https://api.openai.com/v1/realtime/calls/${callId}/accept`;
  const headers = makeHeaders("application/json");
  const body = JSON.stringify(makeSession());

rtc.post("/", async (c) => {
  // Create the call.
  const url = "https://api.openai.com/v1/realtime/calls";
  const headers = makeHeaders();
  const fd = new FormData();

# hello-realtime
Hello Realtime is a complete OpenAI Realtime application that supports both WebRTC and SIP users.
You can access the app via WebRTC at https://hello-realtime.val.run, or via SIP by calling 425-800-0042.
This demo shows off the new SIP API, the new all-in-one WebRTC API, and the new server-side websocket interface.
If you remix the app, you'll just need to pop in your own OPENAI_API_KEY (from platform.openai.com), and also the OPENAI_SIGNING_SECRET if you want to use the SIP interface.

observer.post("/:callId", async (c) => {
  const callId = c.req.param("callId");
  const url = `wss://api.openai.com/v1/realtime?call_id=${callId}`;
  const ws = new WebSocket(url, { headers: makeHeaders() });
  ws.on("open", () => {

    <meta charset="utf-8" />
    <meta name="viewport" content="width=device-width, initial-scale=1" />
    <title>OpenAI Realtime API Voice Agent</title>
    <style>
      :root {