// apps/admin/src/app/(admin)/pipelines/page.tsx // D3RO Voice — AI & Voice Pipeline Intelligence Console (Phase 1~15.5 SSOT) import { Box, Typography, Button } from '@mui/material' import { fetchServerStats } from '@/lib/api-server' import { C, FONT_SANS, FONT_MONO, panelSx, statusBadgeSx } from '@/lib/console-theme' import { StatRing, TactileBadge, DoubleBezelCard } from '@d3ro/ui/components/ds' export default async function PipelinesPage(): Promise { const stats = await fetchServerStats() if (!stats.pipelines) { return ( Pipeline telemetry unavailable The backend does not currently expose measured pipeline telemetry. No simulated engine, latency, accuracy, or capacity values are shown. ) } const { whisper, ollama, realtimeVoice, ragVector, meetingIntelligence } = stats.pipelines return ( <> {/* Header Bar */} AI & Voice Pipeline Matrix ORCHESTRATOR ONLINE Local Whisper, Ollama v0.32.1, GPT-Realtime 2.1, Vector RAG, Diarization {/* Main Grid: 5 Core Pipeline Domains */} {/* Top Split: Whisper STT & Ollama LLM */} {/* Whisper STT Card */} Faster-Whisper STT Sidecar {whisper.activeModel} DEFAULT STT AVG LATENCY {whisper.avgLatencyMs}ms SPEEDUP FACTOR {whisper.speedupFactor} GPU VRAM {whisper.gpuVramUsage} Dual-Condition Flush: Audio buffer flushes parallel with model warm-up.
Realtime Streaming: {whisper.partialStreamingFps} fps interim partial transcription in RecordingTip popup.
Total Today: {whisper.totalTranscriptionsToday.toLocaleString()} voice transcriptions processed.
{/* Bundled Ollama LLM Card */} Bundled Ollama Runtime {ollama.version} (119MB Pruned) LOCAL LLM THROUGHPUT {ollama.tokensPerSecond} tok/s CONTEXT LIMIT {ollama.activeContextLimit} VRAM OCCUPANCY {ollama.vramAllocated} Loaded Models: {ollama.loadedModels.join(', ')}
NDJSON Streaming: Zero-latency token streaming for Auto Polish & AI Chat.
Active Local Sessions: {ollama.activeSessions} concurrent local inference threads.
{/* Middle Split: Realtime Live Voice & Vector RAG */} {/* GPT-Realtime 2.1 Voice Engine */} GPT-Realtime 2.1 Live Engine {realtimeVoice.backend} PRO+ PREMIUM LIVE STREAMS {realtimeVoice.activeStreams} Active AUDIO RTT {realtimeVoice.avgAudioRttMs}ms LOCAL FALLBACK {realtimeVoice.localFallbackRate} Ultra-low Latency: Full-duplex bidirectional voice-to-voice stream.
Auto Failover: Drops gracefully to Local STT + Ollama + TTS if network drops.
Stream Uptime: {realtimeVoice.streamUptime}% across all regional connections.
{/* SQLite Vector RAG Knowledge Base */} SQLite Vector RAG Engine {ragVector.embeddingModel} KNOWLEDGE BASE INDEXED DOCS {ragVector.indexedDocuments.toLocaleString()} VECTOR CHUNKS {ragVector.totalVectorChunks.toLocaleString()} SEARCH HIT RATE {ragVector.topHitRatePercent}% Local Vector Store: Zero-cloud-leakage SQLite embeddings with cosine similarity.
Average Search Latency: {ragVector.avgSearchLatencyMs}ms per 512-dim query.
Semantic Q&A: Grounded context injection into Voice Mode & Meeting Summary.
{/* Bottom Full-Width: Meeting Intelligence & Speaker Diarization (Phase 14~15.5) */} Meeting Intelligence & Speaker Diarization (Phase 14~15.5) {meetingIntelligence.diarizationEngine} PHASE 15.5 COMPLIANT SPEAKER ACCURACY {meetingIntelligence.speakerAccuracyPercent}% Pyannote + LLM Consensus ACTIVE MEETINGS {meetingIntelligence.activeMeetingSessions} Live Realtime Captions & Record TEMPLATES TODAY {meetingIntelligence.templatesGeneratedToday} Docs Summary, Action Items, Jira MINDMAP EXPORTS {meetingIntelligence.mindmapsExported} Maps Interactive Visual Graphs Speaker Attribution Architecture: Hybrid pipeline combining Pyannote 3.1 voiceprint embeddings with LLM conversational speaker inference.
Multi-Document Synthesis (Phase 14.5): Simultaneous generation of Executive Summary, Action Item Checklist, Jira Issue Drafts, and Interactive Markdown Mindmaps.
File Transcription Pipeline: High-speed sequential STT for uploaded MP3/WAV audio recordings up to 4 hours.
) }