# TODO — Ollama

> Last updated: 2026-07-06

---

## High Priority

### Fase 1 — GPU Node (PC-Multimedia) ✅ SELESAI
- [x] CUDA 12.9 sudah tersedia via driver 576.88 (tidak perlu install terpisah)
- [x] Install Ollama v0.31.1 di Windows (`OllamaSetup.exe`)
- [x] Set `OLLAMA_MODELS=T:\ollama\models` + `OLLAMA_HOST=0.0.0.0:11434` via `setx /M`
- [x] Autostart via Task Scheduler (SYSTEM, AtStartup) — script `PC-Multimedia/scripts/install-ollama-task.ps1`
- [x] Firewall rule `OllamaVLAN`: TCP 11434 dari `10.100.1.0/24` only
- [x] Pull `qwen2.5-coder:14b` (9.0 GB) + `nomic-embed-text` (274 MB) → `T:\ollama\models`
- [x] Verified: API `http://10.100.1.14:11434` reachable, VRAM 11.5 GB / 16 GB saat loaded

### Fase 2 — pgvector + Codebase Index ✅ SELESAI
- [x] pgvector Docker di `10.100.1.24:5433` (pgvector 0.8.4 / pg16)
- [x] DB `geonet_project_search` + tabel `code_chunks` + `user_memory` + index HNSW
- [x] Script indexer `Ollama/indexer/indexer.py` — 1.784 chunks dari 1.031 files, 18 repo
- [x] Verified: similarity search akurat (query → FeedbackController.php sebagai top result)

### Fase 3 — Integrasi geonet-console ✅ SELESAI
- [x] `OLLAMA_BASE_URL=http://10.100.1.14:11434` di geonet-console `.env`
- [x] `AiChatService` — embed query → pgvector similarity search → inject context → Ollama chat
- [x] `AiChatController` — `POST /api/v1/ai/chat` + `GET /api/v1/ai/status`
- [x] `config/pgvector.php` — config PGVECTOR_HOST/PORT/DB/USER/PASS
- [x] Verified end-to-end: 5 context chunks retrieved, jawaban akurat tentang route feedback
- [ ] User memory: simpan pola interaksi per user (next iteration)

### VM .26 (tetap jalan sebagai fallback)
- [ ] Tetap running `qwen2.5:7b` + `nomic-embed-text` sebagai fallback CPU inference
- [ ] SMB read-only mount QNAP `10.100.1.10` di VM `.26` (indexer background)

---

## Medium Priority

- [ ] UI chat di Super Apps PWA — window chat terintegrasi
- [ ] User memory / RAG: retrieve pola kebiasaan user saat generate response
- [ ] Deploy agent (Fase 4): tool calling SSH + docker exec untuk bantu deploy
- [ ] Healthcheck / monitoring (Zabbix atau cron `ollama ps`)
- [ ] Optimasi VM `.26`: Ubuntu Desktop → Server atau matikan GUI (hemat RAM)

---

## Low Priority

- [ ] Fine-tuning LoRA untuk template spesifik organisasi (butuh 1000+ contoh data)
- [ ] Evaluasi model `deepseek-coder-v2:16b` untuk coding tasks
- [ ] Backup `/data/ollama/models` ke QNAP
- [ ] Dokumentasi troubleshooting inference timeout
- [ ] Migrasi model ke VM `.26` jika PC-Multimedia tidak bisa 24/7

---

## Selesai (Sesi Ini)

- [x] Audit live VM `10.100.1.26` — Ollama + model verified
- [x] Bootstrap dokumentasi repo `Ollama/` (AI-Agent-Standards)
- [x] Daftar di `Workspace-Context.md`
- [x] Audit hardware PC-Multimedia via DxDiag — Core Ultra 7 265K, 64 GB, RTX 5060 Ti
- [x] Diskusi arsitektur AI onprem: Context Injection → RAG/Memory → Fine-tuning
- [x] Diskusi AI coding assistant + deploy agent on-premise
- [x] **Fase 1 GPU Node PC-Multimedia** — Ollama v0.31.1, GPU inference aktif, verified dari laptop
- [x] **Fase 2 pgvector** — Docker `10.100.1.24:5433`, 1.784 chunks, similarity search verified
- [x] **Fase 3 integrasi** — `POST /api/v1/ai/chat` live, RAG end-to-end verified di geonet-console