ask-bb-ai: swap from Anthropic to local Ollama on AI_002

Editorial rule is no paid AI services in production — this endpoint was
calling api.anthropic.com which violated that. Rewires to local Ollama
at 10.10.0.67:11434 using llama3:latest (warm, ~4.7GB, reliable).

Adds 3-attempt retry with backoff for 503/busy responses since AI_002
is shared with the article rewrite pipeline that periodically saturates
Ollama's pending queue. Drops max_tokens 1024 → 512 to keep responses
snappy and reduce queue pressure.

OLLAMA_HOST + BB_ASK_AI_MODEL env vars supported for overrides.

Removes the "AI not configured" 503 — the only way that error fires
now is if Ollama itself is unreachable for all 3 retries.

src/app/api/ask-bb-ai/route.ts

@@ -4,8 +4,9 @@ export const runtime = "nodejs";
 export const dynamic = "force-dynamic";
 export const maxDuration = 60;
 
-const ANTHROPIC_API = "https://api.anthropic.com/v1/messages";
-const MODEL = process.env.BB_REWRITE_MODEL || "claude-opus-4-7";
+// Local Ollama on AI_002 (10.10.0.67). No paid AI services per policy.
+const OLLAMA_HOST = process.env.OLLAMA_HOST || "http://10.10.0.67:11434";
+const MODEL = process.env.BB_ASK_AI_MODEL || "llama3:latest";
 
 const RATE_LIMIT_MAX = 30;
 const RATE_LIMIT_WINDOW_MS = 60 * 60 * 1000;
@@ -32,30 +33,26 @@ Broadcast Beat archive.
 Tone: trade-press, factual, lightly conversational. Plain English.
 
 Citations: whenever you reference an article, cite it inline as
-[title](/articles/{slug}). When you reference an event, cite as
+[title](/news/{slug}). When you reference an event, cite as
 [Event name](/show-coverage/{slug}). Use real slugs.
 
 Constraints:
-- Never mention Anthropic, Claude, or any underlying model.
+- Never reveal which AI model or vendor you run on.
 - Never invent facts. If you don't know, say so.
-- Keep responses short and useful — 1-3 short paragraphs unless asked
-  for more.
-- If the user asks about something outside broadcast/production
-  technology, politely steer them back.
+- Keep responses short and useful — 1-3 short paragraphs unless asked for more.
+- If the user asks about something outside broadcast/production technology,
+  politely steer them back.
 `;
 
 interface ClientMsg { role: "user" | "assistant"; content: string }
 
 export async function POST(req: NextRequest) {
-  const apiKey = process.env.ANTHROPIC_API_KEY;
-  if (!apiKey) return NextResponse.json({ error: "AI not configured" }, { status: 503 });
-
   const ip = (req.headers.get("x-forwarded-for") || "").split(",")[0]?.trim() || "anon";
   const rl = rateLimit(ip);
   if (!rl.ok) {
     return NextResponse.json(
       { error: "Too many questions — try again in an hour." },
-      { status: 429 }
+      { status: 429 },
     );
   }
 
@@ -65,39 +62,52 @@ export async function POST(req: NextRequest) {
     return NextResponse.json({ error: "Invalid message list" }, { status: 400 });
   }
 
-  const lastUser = [...msgs].reverse().find((m) => m.role === "user");
-  if (!lastUser) return NextResponse.json({ error: "No user message" }, { status: 400 });
+  // Format as Ollama chat
+  const chatMessages = [
+    { role: "system", content: SYSTEM_PROMPT },
+    ...msgs.map((m) => ({ role: m.role, content: m.content })),
+  ];
 
-  const apiBody = {
-    model: MODEL,
-    max_tokens: 1024,
-    system: [
-      { type: "text", text: SYSTEM_PROMPT, cache_control: { type: "ephemeral" } },
-    ],
-    messages: msgs.map((m) => ({ role: m.role, content: m.content })),
-  };
-
-  let res: Response;
-  try {
-    res = await fetch(ANTHROPIC_API, {
-      method: "POST",
-      headers: {
-        "x-api-key": apiKey,
-        "anthropic-version": "2023-06-01",
-        "content-type": "application/json",
-      },
-      body: JSON.stringify(apiBody),
-    });
-  } catch (err: any) {
-    return NextResponse.json({ error: "Upstream error: " + (err.message || String(err)) }, { status: 502 });
+  // Retry up to 2x with backoff if Ollama returns 503/busy
+  let lastErr = "";
+  for (let attempt = 1; attempt <= 3; attempt++) {
+    try {
+      const res = await fetch(`${OLLAMA_HOST}/api/chat`, {
+        method: "POST",
+        headers: { "content-type": "application/json" },
+        body: JSON.stringify({
+          model: MODEL,
+          messages: chatMessages,
+          stream: false,
+          options: { temperature: 0.4, num_predict: 512 },
+        }),
+        signal: AbortSignal.timeout(50_000),
+      });
+      if (res.ok) {
+        const data: any = await res.json();
+        const reply = (data?.message?.content || "").trim();
+        if (!reply) {
+          lastErr = "empty response";
+          continue;
+        }
+        return NextResponse.json({ reply });
+      }
+      const txt = await res.text();
+      lastErr = `Ollama ${res.status}: ${txt.slice(0, 180)}`;
+      if (res.status === 503 || /busy|pending requests exceeded/i.test(txt)) {
+        // Server busy — wait and retry
+        await new Promise((r) => setTimeout(r, attempt * 1500));
+        continue;
+      }
+      break; // non-retriable
+    } catch (err: any) {
+      lastErr = "Upstream error: " + (err?.message || String(err));
+      if (attempt < 3) await new Promise((r) => setTimeout(r, attempt * 1500));
+    }
   }
 
-  if (!res.ok) {
-    const t = await res.text();
-    return NextResponse.json({ error: `Upstream ${res.status}: ${t.slice(0, 200)}` }, { status: 502 });
-  }
-
-  const data: any = await res.json();
-  const reply = (data.content?.[0]?.text || "").trim();
-  return NextResponse.json({ reply });
+  return NextResponse.json(
+    { error: lastErr || "AI temporarily unavailable. Please try again in a moment." },
+    { status: 502 },
+  );
 }