lcbp3/specs/300-others/302-ai-model-revision/quickstart.md

Quickstart: AI Model Revision (ADR-023A)

Feature: 302-ai-model-revision Date: 2026-05-15

---

Prerequisites

1. Ollama on Desk-5439 (AI Inference Engine)

Requirements:

• OS: Windows 10/11 หรือ Linux (Desk-5439)
• GPU: NVIDIA GPU ที่รองรับ CUDA 11.8+ (VRAM ≥ 6GB แนะนำ)
• Ollama Version: ≥ 0.5.0
• Models: gemma4:e2b (Q4_K_M quantization) + nomic-embed-text

Verification Steps:

# 1. Check Ollama installation
ollama --version
# Expected: ollama version is 0.5.0 or higher

# 2. Verify GPU access
nvidia-smi
# Expected: NVIDIA GPU detected with CUDA 11.8+

# 3. List installed models
ollama list
# Expected output:
# NAME                    ID              SIZE      MODIFIED
# gemma4:e2b              <hash>          2.4 GB    <timestamp>
# nomic-embed-text        <hash>          274 MB    <timestamp>

# 4. Test model inference (quick test)
ollama run gemma4:e2b "Hello, test response"
# Expected: Gemma4 generates text response within 5-10s

# 5. Test embedding model
ollama run nomic-embed-text "test text"
# Expected: Returns 768-dimensional vector

Installation Commands (if missing):

# Install Ollama (Windows)
# Download from: https://ollama.com/download/windows

# Pull models
ollama pull gemma4:e2b
ollama pull nomic-embed-text

# Verify VRAM usage during inference
nvidia-smi --query-gpu=memory.used --format=csv,noheader
# Expected: < 5120 MB (5GB threshold per SC-003)

Troubleshooting:

• Ollama not responding: Check if Ollama service is running (services.msc on Windows)
• GPU not detected: Verify NVIDIA driver and CUDA installation
• Model pull fails: Check internet connection and proxy settings
• VRAM overflow: Reduce concurrent jobs (BullMQ concurrency=1 enforced)

---

2. Qdrant Vector Database (Semantic Search)

Requirements:

• Version: Qdrant ≥ 1.7.0
• Storage: ≥ 10GB free space (for ~20,000 docs × 768-dim vectors)
• Network: Accessible from backend container (http://QDRANT_HOST:6333)
• Collection: lcbp3_documents (auto-created on first use)

Verification Steps:

# 1. Check Qdrant health
curl http://192.168.10.XX:6333/health
# Expected: {"status":"ok"}

# 2. Check collections (if already created)
curl http://192.168.10.XX:6333/collections
# Expected: {"result":{"collections":[...]}} or empty array

# 3. Verify Qdrant metrics (optional)
curl http://192.168.10.XX:6333/metrics
# Expected: Prometheus metrics output

Installation Commands (Docker - recommended):

# Run Qdrant container (QNAP or ASUSTOR)
docker run -d \
  --name qdrant \
  -p 6333:6333 \
  -p 6334:6334 \
  -v $(pwd)/qdrant_storage:/qdrant/storage \
  qdrant/qdrant:v1.7.0

# Or use docker-compose (add to existing compose file)
# See: specs/04-Infrastructure-OPS/04-00-docker-compose/

Troubleshooting:

• Connection refused: Check if Qdrant container is running (docker ps)
• Port conflict: Ensure port 6333 is not used by other services
• Storage errors: Verify disk space and permissions on mounted volume
• Network issues: Check firewall rules allow traffic on port 6333

---

3. Redis 7 (BullMQ Queue Backend)

Requirements:

• Version: Redis ≥ 7.0.0
• Persistence: AOF enabled (recommended for job durability)
• Memory: ≥ 2GB (for job queue storage)
• Network: Accessible from backend container (redis://REDIS_HOST:6379)

Verification Steps:

# 1. Check Redis version
redis-cli -h 192.168.10.XX -p 6379 INFO server | grep redis_version
# Expected: redis_version=7.x.x

# 2. Check Redis health
redis-cli -h 192.168.10.XX -p 6379 PING
# Expected: PONG

# 3. Check existing queues (if any)
redis-cli -h 192.168.10.XX -p 6379 KEYS "bull:*"
# Expected: May show existing BullMQ queues from other modules

# 4. Check memory usage
redis-cli -h 192.168.10.XX -p 6379 INFO memory | grep used_memory_human
# Expected: used_memory_human:<value> (monitor for growth)

Configuration Notes:

• Use existing Redis instance from BullMQ setup (ADR-008)
• No additional configuration needed for 2-queue setup
• Queue names: ai-realtime and ai-batch (auto-created by BullMQ)

Troubleshooting:

• Connection timeout: Check Redis service status and network connectivity
• Memory overflow: Monitor Redis memory usage, enable maxmemory policy if needed
• Queue not created: Check BullMQ configuration in backend/src/config/bullmq.config.ts

---

4. Backend Dependencies (NestJS + AI Libraries)

Requirements:

• Node.js: ≥ 18.x (LTS)
• npm/pnpm: Latest version
• TypeScript: ≥ 5.x

Required Packages:

# Navigate to backend directory
cd backend

# Check installed packages
npm ls @qdrant/js-client-rest
# Expected: @qdrant/js-client-rest@1.x.x

# Check other AI-related packages
npm ls bullmq ioredis
# Expected: bullmq@5.x.x, ioredis@5.x.x

# Install missing packages (if any)
npm install @qdrant/js-client-rest
npm install bullmq ioredis --save

Verification Steps:

# 1. Check package.json
cat package.json | grep -A 5 "dependencies"
# Expected: Should include @qdrant/js-client-rest, bullmq, ioredis

# 2. Verify TypeScript compilation
npm run build
# Expected: No compilation errors

# 3. Check environment variables
cat .env | grep -E "OLLAMA|QDRANT|OCR|REDIS"
# Expected: All AI-related variables set (see Environment Variables section)

Troubleshooting:

• Package not found: Run npm install to restore dependencies
• TypeScript errors: Check tsconfig.json and type definitions
• Env variables missing: Copy from .env.example and configure

---

5. PaddleOCR Sidecar (OCR Engine)

Requirements:

• Python: ≥ 3.8
• PaddlePaddle: ≥ 2.5.0
• PaddleOCR: ≥ 2.7.0
• API Server: HTTP endpoint on port 8765 (default)

Verification Steps:

# 1. Check PaddleOCR API health
curl http://localhost:8765/health
# Expected: {"status":"ok"} or similar

# 2. Test OCR on sample image
curl -X POST http://localhost:8765/ocr \
  -F "image=@sample.png" \
  -F "lang=th"
# Expected: JSON with extracted text

# 3. Check OCR character threshold setting
# (Backend env var: OCR_CHAR_THRESHOLD=100)
# This determines when to use PaddleOCR vs PyMuPDF

Installation Commands (if not already set up):

# Install PaddleOCR (Python)
pip install paddleocr paddlepaddle

# Run OCR API server (example)
python -m paddleocr.serve --port 8765

Troubleshooting:

• API not responding: Check if PaddleOCR server is running
• OCR fails on Thai text: Verify Thai language model is installed
• Slow performance: Consider GPU acceleration for PaddleOCR

---

6. Network Configuration & Firewall

Requirements:

• Desk-5439 (Ollama): Port 11434 accessible from backend
• Qdrant Server: Port 6333 accessible from backend
• Redis Server: Port 6379 accessible from backend
• PaddleOCR: Port 8765 accessible from backend (if sidecar on same host)

Verification Steps:

# Test connectivity from backend container
# (Run these commands inside backend container or from backend host)

# 1. Test Ollama
curl http://192.168.10.XX:11434/api/tags
# Expected: JSON with model list

# 2. Test Qdrant
curl http://192.168.10.XX:6333/health
# Expected: {"status":"ok"}

# 3. Test Redis
redis-cli -h 192.168.10.XX -p 6379 PING
# Expected: PONG

# 4. Test PaddleOCR (if on separate host)
curl http://192.168.10.XX:8765/health
# Expected: {"status":"ok"}

Troubleshooting:

• Connection timeout: Check firewall rules on target hosts
• DNS resolution issues: Use IP addresses instead of hostnames
• VPN/Network segmentation: Ensure backend can reach AI infrastructure network

---

7. GPU Resource Monitoring (Critical for SC-003)

Requirements:

• VRAM Limit: ≤ 5GB peak (per SC-003)
• Concurrency: 1 job per queue (enforced by BullMQ)

Verification Commands:

# Monitor VRAM in real-time
watch -n 1 nvidia-smi --query-gpu=memory.used --format=csv,noheader

# Check GPU utilization during job execution
nvidia-smi dmon -s u

# Log VRAM usage to file (for analysis)
nvidia-smi --query-gpu=timestamp,memory.used,utilization.gpu \
  --format=csv -l 1 > gpu_usage.log

Expected Behavior:

• ai-batch job: VRAM peaks at ~2.5GB (gemma4:e2b Q4_K_M)
• ai-realtime job: VRAM peaks at ~2.5GB (same model)
• No concurrent jobs: ai-batch pauses when ai-realtime active (GPU protection)

Troubleshooting:

• VRAM overflow (>5GB): Reduce model quantization or increase GPU memory
• GPU contention: Verify BullMQ concurrency=1 enforcement
• Slow inference: Check GPU utilization, consider faster model quantization

---

Environment Variables (เพิ่มใน backend/.env)

# AI Infrastructure
OLLAMA_HOST=http://192.168.10.XX:11434
OLLAMA_MODEL_MAIN=gemma4:e2b
OLLAMA_MODEL_EMBED=nomic-embed-text

# Qdrant
QDRANT_HOST=http://192.168.10.XX:6333
QDRANT_COLLECTION=lcbp3_documents

# OCR
OCR_CHAR_THRESHOLD=100
OCR_API_URL=http://localhost:8765   # PaddleOCR sidecar

# BullMQ Queue Names (for reference — hardcoded in code)
# ai-realtime: RAG Q&A, AI Suggest
# ai-batch: OCR, Extract, Embed

---

Verification Scenarios (สำหรับ QA / UAT)

Scenario 1: Digital PDF Classification (Fast Path)

# 1. Upload a digital PDF via RFA form
# 2. Monitor BullMQ dashboard (if Bull Board installed)
# Expected: ai-realtime queue → ai-suggest job → completed within 30s
# Expected: ai-batch queue → embed-document job → completed in background
curl -H "Authorization: Bearer $TOKEN" http://localhost:3001/api/ai/jobs/status/$JOB_ID

Scenario 2: Scanned PDF (OCR Path)

# 1. Upload a scanned (image-only) PDF
# Expected: ai-batch queue → ocr job first → then ai-suggest
# OCR detection: extracted_chars < 100 per page triggers slow path

Scenario 3: RAG Query (Multi-tenancy Check)

# 1. Embed 2 docs in Project A, 1 doc in Project B
# 2. Query from Project A context
# Expected: results contain ONLY Project A docs
curl -X POST http://localhost:3001/api/ai/rag/query \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"projectPublicId":"<project-a-uuid>","question":"ค้นหา shop drawing หมายเลข SD-001"}'

Scenario 4: Legacy Migration (Admin Only)

# 1. Trigger via n8n (or direct API for testing):
curl -X POST http://localhost:3001/api/ai/migration/queue \
  -H "Authorization: Bearer $ADMIN_TOKEN" \
  -H "Idempotency-Key: SD-001-2026:batch-001" \
  -H "Content-Type: application/json" \
  -d '{"batchId":"batch-001","filename":"SD-001.pdf","tempPath":"/uploads/temp/SD-001.pdf"}'
# 2. Check migration_review_queue: status = PENDING
# 3. Admin Approve from UI → status = IMPORTED

Scenario 5: Typhoon Removal Verification

# After implementation, run:
grep -r "typhoon" backend/src --include="*.ts"
# Expected: NO results (file should be deleted)

Scenario 6: GPU Overload Prevention + VRAM Verification

# 1. While ai-batch job is running, submit ai-realtime job
# Expected: ai-batch pauses; ai-realtime job completes; ai-batch resumes
# Observable via BullMQ dashboard or job status API

# 2. Measure VRAM peak during job run (verify SC-003):
nvidia-smi --query-gpu=memory.used --format=csv,noheader
# Expected: value < 5120 MB (5GB threshold per SC-003)
# Repeat during both ai-batch and ai-realtime jobs to verify peak

---

Key Files to Modify (Priority Order)

Priority	File	Change
🔴 CRITICAL	backend/src/modules/ai/rag/typhoon.service.ts	DELETE
🔴 CRITICAL	backend/src/config/bullmq.config.ts	Add ai-batch queue
🔴 CRITICAL	specs/03-Data-and-Storage/deltas/14-add-migration-review-queue.sql	CREATE
🟡 HIGH	backend/src/modules/ai/services/qdrant.service.ts	Enforce projectPublicId param
🟡 HIGH	backend/src/modules/ai/processors/ai-batch.processor.ts	NEW — replace old processor
🟡 HIGH	backend/src/modules/ai/services/ocr.service.ts	NEW — auto-detect routing
🟢 NORMAL	frontend/app/(dashboard)/ai-staging/page.tsx	Add Migration Queue tab