From 42f7298d85ac7eacdf2428060702be3071d6cf5a Mon Sep 17 00:00:00 2001
From: Remi <73385395+itsrubberduck@users.noreply.github.com>
Date: Sat, 20 Sep 2025 09:59:18 +0200
Subject: [PATCH] feat: add LLM latency check endpoint

---
 server/api/llm/latency.get.ts | 68 +++++++++++++++++++++++++++++++++++
 1 file changed, 68 insertions(+)
 create mode 100644 server/api/llm/latency.get.ts

diff --git a/server/api/llm/latency.get.ts b/server/api/llm/latency.get.ts
new file mode 100644
index 0000000..97ecfd9
--- /dev/null
+++ b/server/api/llm/latency.get.ts
@@ -0,0 +1,68 @@
+// server/api/llm/latency.get.ts
+import { createError } from 'h3'
+import { getOpenAIClient } from '../../utils/openai'
+import { getServerRuntimeConfig } from '../../utils/runtimeConfig'
+
+const CLEARANCE_MESSAGE = [
+    'ATC clearance:',
+    'Lufthansa four seven eight, cleared to Frankfurt via the NORDA one alpha departure,',
+    'climb and maintain five thousand, expect flight level three two zero ten minutes after departure,',
+    'departure frequency one two zero decimal eight, squawk four two one three.'
+].join(' ')
+
+const READBACK_MESSAGE = [
+    'Pilot readback:',
+    'Lufthansa four seven eight is cleared to Frankfurt via NORDA one alpha,',
+    'climb maintain five thousand, expect flight level three two zero in ten,',
+    'departure one two zero decimal eight, squawk four two one three.'
+].join(' ')
+
+const SYSTEM_PROMPT = [
+    'You verify IFR clearance readbacks for aviation radio communication.',
+    'Answer with a single digit only:',
+    '1 when the readback correctly repeats the clearance,',
+    '0 when part of the readback is wrong or missing,',
+    '2 when the transmission is not a valid readback or unrelated to the clearance.'
+].join(' ')
+
+export default defineEventHandler(async () => {
+    const client = getOpenAIClient()
+    const { llmModel } = getServerRuntimeConfig()
+    const model = llmModel || 'chatgpt-5-nano'
+
+    const started = Date.now()
+
+    try {
+        const response = await client.chat.completions.create({
+            model,
+            temperature: 0,
+            max_tokens: 1,
+            messages: [
+                { role: 'system', content: SYSTEM_PROMPT },
+                {
+                    role: 'user',
+                    content: [CLEARANCE_MESSAGE, READBACK_MESSAGE].join('\n')
+                }
+            ]
+        })
+
+        const raw = response.choices?.[0]?.message?.content?.trim() || ''
+        const parsed = Number.parseInt(raw, 10)
+        const validResult = Number.isInteger(parsed) && parsed >= 0 && parsed <= 2 ? parsed : null
+        const latencyMs = Date.now() - started
+
+        return {
+            result: validResult,
+            raw,
+            latency_ms: latencyMs
+        }
+    } catch (error: any) {
+        const latencyMs = Date.now() - started
+        console.error('[LLM] Latency check failed:', error)
+        throw createError({
+            statusCode: error?.status ?? 500,
+            statusMessage: error?.message ?? 'Latency check failed',
+            data: { latency_ms: latencyMs }
+        })
+    }
+})