Wire new routes; identity, summarize-url, dashboard, admin
This commit is contained in:
Keysat
+118
-57
@@ -28,8 +28,13 @@
|
||||
|
||||
import express from "express";
|
||||
import multer from "multer";
|
||||
import { resolveLicense } from "../keysat-client.js";
|
||||
import { getOrCreateRow, planBackend, commitCredit } from "../credits.js";
|
||||
import { resolveIdentity, identityTier } from "../identity.js";
|
||||
import {
|
||||
getOrCreateRow,
|
||||
planBackend,
|
||||
commitCredit,
|
||||
licenseFingerprint,
|
||||
} from "../credits.js";
|
||||
import { lookupJob, markJobCharged, refundJob } from "../job-credits.js";
|
||||
import { getConfigSnapshot, getTierQuotas } from "../config.js";
|
||||
import { createGeminiBackend } from "../backends/gemini.js";
|
||||
@@ -37,6 +42,9 @@ import { createHardwareBackend } from "../backends/hardware.js";
|
||||
import { envelope, errorEnvelope } from "./envelope.js";
|
||||
import { recordCall } from "../audit-log.js";
|
||||
import { calcGeminiCost } from "../pricing.js";
|
||||
import { getAudioDurationSecondsFromBuffer } from "../audio-meta.js";
|
||||
import { resolveHardwareConfig } from "../hardware-config.js";
|
||||
import { reportHealthEvent } from "../spark-control-events.js";
|
||||
|
||||
const upload = multer({
|
||||
storage: multer.memoryStorage(),
|
||||
@@ -48,67 +56,93 @@ export function transcribeRouter() {
|
||||
|
||||
router.post("/transcribe", upload.single("audio"), async (req, res) => {
|
||||
const t0 = Date.now();
|
||||
const installId = req.header("X-Recap-Install-Id");
|
||||
const jobId = req.header("X-Recap-Job-Id") || null;
|
||||
const auth = req.header("Authorization");
|
||||
|
||||
if (!installId) {
|
||||
let identity;
|
||||
try {
|
||||
identity = await resolveIdentity(req);
|
||||
} catch (err) {
|
||||
const e = await errorEnvelope({
|
||||
error: err?.message || "auth_error",
|
||||
statusHint: err?.status || 401,
|
||||
});
|
||||
return res.status(e.statusHint || 401).json(e.body);
|
||||
}
|
||||
if (identity.kind === "license" && !identity.installId) {
|
||||
const e = await errorEnvelope({
|
||||
error: "missing X-Recap-Install-Id header",
|
||||
statusHint: 400,
|
||||
});
|
||||
return res.status(400).json(e.body);
|
||||
}
|
||||
const { creditKey, installId, license } = identity;
|
||||
if (!req.file) {
|
||||
const e = await errorEnvelope({ error: "missing audio file", installId, statusHint: 400 });
|
||||
const e = await errorEnvelope({ error: "missing audio file", creditKey, installId, statusHint: 400 });
|
||||
return res.status(400).json(e.body);
|
||||
}
|
||||
|
||||
const license = await resolveLicense(auth);
|
||||
const tier = license.tier;
|
||||
|
||||
const row = await getOrCreateRow(installId);
|
||||
const row = await getOrCreateRow({ creditKey, installId, license });
|
||||
const tier = identityTier(identity, row);
|
||||
row.tier_snapshot = tier;
|
||||
const licenseFp = identity.kind === "cloud" ? null : licenseFingerprint(license);
|
||||
const auditInstall = installId || identity.userId || null;
|
||||
|
||||
let reusedJob = false;
|
||||
let chosenBackend = null;
|
||||
const existingJob = lookupJob(installId, jobId);
|
||||
if (existingJob) {
|
||||
reusedJob = true;
|
||||
chosenBackend = existingJob.backend;
|
||||
} else {
|
||||
const cfg = await getConfigSnapshot();
|
||||
const hasHardware = !!cfg.relay_parakeet_base_url;
|
||||
const quota = await getTierQuotas();
|
||||
const preference =
|
||||
cfg.relay_transcribe_backend_preference || "gemini_first";
|
||||
const plan = planBackend(row, quota, { hasHardware, preference });
|
||||
if (!plan.allowed) {
|
||||
await recordCall({
|
||||
install_id: installId,
|
||||
tier,
|
||||
pipeline: "transcribe",
|
||||
backend: null,
|
||||
model: null,
|
||||
status: "refused",
|
||||
credit_charged: 0,
|
||||
duration_ms: Date.now() - t0,
|
||||
cost_usd: 0,
|
||||
job_id: jobId,
|
||||
error: plan.reason,
|
||||
});
|
||||
const e = await errorEnvelope({
|
||||
error: plan.reason,
|
||||
installId,
|
||||
tier,
|
||||
statusHint: 402,
|
||||
});
|
||||
return res.status(402).json(e.body);
|
||||
}
|
||||
chosenBackend = plan.backend;
|
||||
}
|
||||
// Probe audio duration BEFORE the backend call so we can record
|
||||
// it on every audit row (success and error alike). Used by the
|
||||
// dashboard to normalize wall-clock time to "ms per minute of
|
||||
// audio" — a backend-agnostic speed benchmark.
|
||||
const audioSeconds = await getAudioDurationSecondsFromBuffer(
|
||||
req.file?.buffer
|
||||
);
|
||||
|
||||
// Billing vs. routing are decoupled — see analyze.js for the
|
||||
// full reasoning. Look up job to decide whether to charge a
|
||||
// credit, but always run planBackend fresh so transcribe's
|
||||
// routing decision respects relay_transcribe_backend_preference.
|
||||
const reusedJob = !!lookupJob({ creditKey, installId, license, jobId });
|
||||
const cfg = await getConfigSnapshot();
|
||||
const hw = await resolveHardwareConfig(cfg);
|
||||
// Operator-only diagnostic — see the matching comment in
|
||||
// summarize-url.js for the full reasoning. We don't 503 here on
|
||||
// blocked_reason because doing so pre-empts planBackend and
|
||||
// surfaces operator-internal wording to clients even when
|
||||
// Gemini was the configured preference.
|
||||
if (hw.transcribe.blocked_reason) {
|
||||
console.warn(
|
||||
`[transcribe] hardware transcribe currently blocked (planBackend will route to Gemini if available): ${hw.transcribe.blocked_reason}`,
|
||||
);
|
||||
}
|
||||
const hasHardware = !!hw.transcribe.url;
|
||||
const quota = await getTierQuotas();
|
||||
const preference =
|
||||
cfg.relay_transcribe_backend_preference || "gemini_first";
|
||||
const plan = planBackend(row, quota, { hasHardware, preference });
|
||||
if (!plan.allowed) {
|
||||
await recordCall({
|
||||
install_id: auditInstall,
|
||||
license_fingerprint: licenseFp,
|
||||
tier,
|
||||
pipeline: "transcribe",
|
||||
backend: null,
|
||||
model: null,
|
||||
status: "refused",
|
||||
credit_charged: 0,
|
||||
duration_ms: Date.now() - t0,
|
||||
audio_seconds: audioSeconds,
|
||||
cost_usd: 0,
|
||||
job_id: jobId,
|
||||
error: plan.reason,
|
||||
});
|
||||
const e = await errorEnvelope({
|
||||
error: plan.reason,
|
||||
creditKey,
|
||||
installId,
|
||||
tier,
|
||||
statusHint: 402,
|
||||
});
|
||||
return res.status(402).json(e.body);
|
||||
}
|
||||
const chosenBackend = plan.backend;
|
||||
let result;
|
||||
try {
|
||||
if (chosenBackend === "gemini") {
|
||||
@@ -116,6 +150,8 @@ export function transcribeRouter() {
|
||||
apiKey: cfg.relay_gemini_api_key,
|
||||
transcriptionModel: cfg.relay_gemini_transcription_model,
|
||||
analysisModel: cfg.relay_gemini_analysis_model,
|
||||
txChunkSeconds: (cfg.relay_gemini_tx_chunk_minutes || 30) * 60,
|
||||
txConcurrency: cfg.relay_gemini_tx_concurrency || 12,
|
||||
});
|
||||
result = await backend.transcribeAudio({
|
||||
audio: req.file.buffer,
|
||||
@@ -128,10 +164,19 @@ export function transcribeRouter() {
|
||||
});
|
||||
} else {
|
||||
const backend = createHardwareBackend({
|
||||
parakeetBaseURL: cfg.relay_parakeet_base_url,
|
||||
gemmaBaseURL: cfg.relay_gemma_base_url,
|
||||
parakeetModel: cfg.relay_parakeet_model,
|
||||
gemmaModel: cfg.relay_gemma_model,
|
||||
parakeetBaseURL: hw.transcribe.url || "",
|
||||
gemmaBaseURL: hw.analyze.url || "",
|
||||
sparkControlBaseURL: hw.sparkBase || "",
|
||||
parakeetModel: hw.transcribe.model || "",
|
||||
gemmaModel: hw.analyze.model || "",
|
||||
txChunkSeconds: (cfg.relay_hardware_tx_chunk_minutes || 5) * 60,
|
||||
txChunkOverlapSeconds: cfg.relay_hardware_tx_chunk_overlap_seconds ?? 30,
|
||||
diarizationEnabled: !!cfg.relay_hardware_diarization_enabled,
|
||||
clusterThresholdPct: cfg.relay_hardware_voice_clustering_threshold ?? 70,
|
||||
anchorMinSpeakingSec: cfg.relay_hardware_anchor_min_speaking_sec ?? 30,
|
||||
smallClusterMaxSpeakingSec: cfg.relay_hardware_small_cluster_max_speaking_sec ?? 15,
|
||||
uncertainMarginPct: cfg.relay_hardware_uncertain_margin_pct ?? 10,
|
||||
txConcurrency: cfg.relay_hardware_tx_concurrency || 4,
|
||||
});
|
||||
result = await backend.transcribeAudio({
|
||||
audio: req.file.buffer,
|
||||
@@ -140,25 +185,38 @@ export function transcribeRouter() {
|
||||
});
|
||||
}
|
||||
} catch (err) {
|
||||
if (reusedJob) refundJob(installId, jobId);
|
||||
if (reusedJob) await refundJob({ creditKey, installId, license, jobId });
|
||||
console.error(`[relay/transcribe] backend error: ${err?.message}`);
|
||||
// Fire-and-forget health report for hardware-served calls;
|
||||
// Gemini failures are a separate observability surface.
|
||||
if (chosenBackend === "hardware") {
|
||||
reportHealthEvent({
|
||||
service: "parakeet",
|
||||
ok: false,
|
||||
error: (err?.message || String(err)).slice(0, 280),
|
||||
ms: Date.now() - t0,
|
||||
});
|
||||
}
|
||||
await recordCall({
|
||||
install_id: installId,
|
||||
install_id: auditInstall,
|
||||
license_fingerprint: licenseFp,
|
||||
tier,
|
||||
pipeline: "transcribe",
|
||||
backend: chosenBackend,
|
||||
model: chosenBackend === "gemini"
|
||||
? cfg.relay_gemini_transcription_model
|
||||
: cfg.relay_parakeet_model,
|
||||
: hw.transcribe.model || "(auto)",
|
||||
status: "error",
|
||||
credit_charged: 0,
|
||||
duration_ms: Date.now() - t0,
|
||||
audio_seconds: audioSeconds,
|
||||
cost_usd: 0,
|
||||
job_id: jobId,
|
||||
error: (err?.message || String(err)).slice(0, 200),
|
||||
});
|
||||
const e = await errorEnvelope({
|
||||
error: err?.message || "backend_error",
|
||||
creditKey,
|
||||
installId,
|
||||
tier,
|
||||
statusHint: err?.status || 502,
|
||||
@@ -168,8 +226,8 @@ export function transcribeRouter() {
|
||||
|
||||
let creditCharged = 0;
|
||||
if (!reusedJob) {
|
||||
await commitCredit(installId, { backend: chosenBackend, tier });
|
||||
markJobCharged(installId, jobId, { backend: chosenBackend, tier });
|
||||
await commitCredit({ creditKey, installId, license, backend: chosenBackend, tier });
|
||||
await markJobCharged({ creditKey, installId, license, jobId, backend: chosenBackend, tier });
|
||||
creditCharged = 1;
|
||||
}
|
||||
|
||||
@@ -188,6 +246,7 @@ export function transcribeRouter() {
|
||||
};
|
||||
await recordCall({
|
||||
install_id: installId,
|
||||
license_fingerprint: licenseFp,
|
||||
tier,
|
||||
pipeline: "transcribe",
|
||||
backend: chosenBackend,
|
||||
@@ -195,11 +254,13 @@ export function transcribeRouter() {
|
||||
status: "success",
|
||||
credit_charged: creditCharged,
|
||||
duration_ms: Date.now() - t0,
|
||||
audio_seconds: audioSeconds,
|
||||
job_id: jobId,
|
||||
attempts: result?.attempts || null,
|
||||
...costDetails,
|
||||
});
|
||||
|
||||
const body = await envelope({ result, installId, tier, creditCharged });
|
||||
const body = await envelope({ result, creditKey, installId, license, tier, creditCharged });
|
||||
res.json(body);
|
||||
});
|
||||
|
||||
|
||||
Reference in New Issue
Block a user