np-dms/lcbp3

Fork 0

Files

T

admin 3274dede7a

CI / CD Pipeline / build (push) Failing after 4m28s

Details

CI / CD Pipeline / deploy (push) Has been skipped

Details

690603:2041 ADR-034-134 #01

2026-06-03 20:41:42 +07:00

3.8 KiB

Raw Blame History

// File: specs/100-Infrastructures/134-ai-model-change/data-model.md // Change Log: // - 2026-06-03: Data model for Thai-Optimized AI Model Stack

Data Model: Thai-Optimized AI Model Stack

Database Schema Changes

ไม่มี new tables หรือ column changes — ADR-009 compliant; ไม่มี TypeORM migration

Optional: ai_available_models Seed Update

ตาราง ai_available_models (จาก ADR-027 Session 6) ควร update เพื่อความสอดคล้อง:

Update existing main model record:

UPDATE ai_available_models
SET model_name = 'typhoon2.5-np-dms:latest',
    display_name = 'Typhoon 2.5 NP-DMS (Thai)',
    updated_at = NOW()
WHERE model_type = 'main' AND is_active = 1;

Insert OCR model record (if column model_type supports 'ocr'):

INSERT INTO ai_available_models (model_name, display_name, model_type, is_active, created_at, updated_at)
VALUES ('typhoon-np-dms-ocr:latest', 'Typhoon OCR 3B (Thai)', 'ocr', 1, NOW(), NOW())
ON DUPLICATE KEY UPDATE display_name = VALUES(display_name), updated_at = NOW();

Note

: ตรวจสอบ schema จริงใน lcbp3-v1.9.0-schema-02-tables.sql ก่อน run delta — column names และ model_type ENUM อาจแตกต่าง

Code Configuration Model

AiSettingsService Constants

Constant	เดิม	ใหม่
`DEFAULT_MODEL`	`'gemma4:e2b'`	`'typhoon2.5-np-dms:latest'`
`OCR_MODEL` (เพิ่มใหม่)	—	`'typhoon-np-dms-ocr:latest'`

File: backend/src/modules/ai/services/ai-settings.service.ts

Custom Ollama Model Configurations

typhoon2.5-np-dms (Main Model)

Property	Value
Custom Name	`typhoon2.5-np-dms:latest`
Base Model	`scb10x/typhoon2.5-qwen3-4b:latest`
Size	~2.5GB VRAM
keep_alive	Indefinite (`-1`) — standby ตลอด
num_ctx	8192
num_predict	2048
temperature	0.1
top_p	0.85
repeat_penalty	1.15
Role	Extraction, RAG Q&A, AI Suggestion, OCR Post-processing
Modelfile Path	`specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/typhoon2.5-np-dms.model.md`

typhoon-np-dms-ocr (OCR Model)

Property	Value
Custom Name	`typhoon-np-dms-ocr:latest`
Base Model	`scb10x/typhoon-ocr1.5-3b:latest`
Size	~3.2GB VRAM
keep_alive	`0` — auto-unload immediately after job
num_ctx	8192
num_predict	4096
temperature	0.1
top_p	0.1
repeat_penalty	1.1
Role	Thai OCR extraction from PDF images
Modelfile Path	`specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/typhoon-np-dms-ocr.model.md`

VRAM Budget Analysis

State	Models in VRAM	Estimated VRAM Usage
Idle	typhoon2.5-np-dms (standby)	~2.5GB
Main AI job	typhoon2.5-np-dms (active)	~3.5–4GB peak
OCR transition	unloading main → loading OCR	~3.2GB (OCR only)
OCR processing	typhoon-np-dms-ocr (active)	~4–5GB peak
Post-OCR reload	loading main back	~2.5GB
Max peak	OCR model active	~5GB (safe under 8GB limit)

OllamaService New Methods Signatures

/**
 * unloadModel: Force unload model จาก VRAM
 * @param modelName - ชื่อ Ollama model ที่ต้องการ unload
 */
async unloadModel(modelName: string): Promise<void>

/**
 * loadModel: Load / warm model เข้า VRAM
 * @param modelName - ชื่อ Ollama model ที่ต้องการ load
 * @param keepAlive - -1 = indefinite (default สำหรับ main model); 0 = auto-unload
 */
async loadModel(modelName: string, keepAlive?: number | string): Promise<void>

3.8 KiB Raw Blame History Unescape Escape