HANDOFF — 2026-05-02 23:30 — FORENSIC FINAL

🚨 KRITIČNO: Damir bio u pravu — LoRA postoji

Prethodne sesije pogrešno tvrdile "fine-tune nije rađen". Forenzička pretraga otkrila 4 LoRA treninga:

Adapter	Base	Loss train/eval	Accuracy	Status
dabi-budget-lora-v1	Qwen2.5-7B	0.489/0.329	92.7%	✅ PRODUCTION na :8765
dabi-qwen-3b-v1	Qwen2.5-3B	-	-	Final + 3 ckpt
dabi-qlora-qwen25-3b-v1	Qwen2.5-3B	0.515	-	Final + 2 ckpt
dabi-qwen1.5b-v1	Qwen2.5-1.5B	-	-	Final + 2 ckpt

F10_lora_server.py radi 5+ dana neprekidno na portu 8765, OpenAI compatible API, llama-cpp-python CPU mode.

✅ ŠTO JE DOVRŠENO U OVOJ SESIJI

A) Forenzička provjera + memory korekcija

63 active custom services (ne 17)
Hardware Intel i5-13500 (ne AMD)
Već postojeći servisi: rinet-mcp :8810, 3 reranker @8099/8100/8101, F10 LoRA :8765, llm-router

B) Duplikati uklonjeni

❌ moj rinet-reranker :9881 → STOPPED + DISABLED
Orchestrator preusmjeren na pool 8099/8100/8101

C) dabi-budget LoRA integriran

Tier 0 u LLMWaterfall za BUDGET_KEYWORDS
localhost:8765/v1/chat/completions

D) PRIORITY_QA_DB_FIRST bypass

Lookup u dabi.knowledge (sources: nk_rijeka_official, damir_priority_facts) preko FTS to_tsquery
PRIJE intent classification — sprečava SQL precedence problem
escape %% za psycopg2
Returns rag_qa_direct_db source_type, model_used db_priority_lookup

E) CULTURAL_QA_BYPASS + BUDGET_TOTAL_BYPASS

"puta osvojila/prvenstvo/Kup HR" → forces RAG path (ne SQL)
"proracun PGZ" bez "sport" → forces RAG path (ne pgz_sport.proracun)

F) Outdated facts brisani

1× damir_priority_facts (1× Prvenstvo, 6× Kup) — DELETED
5× Qdrant orphans (cultural_v2 + knowledge_v2 + pgz_sport_v1) — DELETED

G) lora-finetune.timer DAILY 03:00

Aktivan, randomized 0-30min
NEXT: 2026-05-03 03:10 CEST
Triggers lora-finetune.service

📊 SMOKE TEST 5/5 PASS

Pitanje	Source	Točan?
Koliko Rijeka prvenstvo?	rag_qa_direct_db	✅ DVA PUTA
Koliko Kupova HR?	rag_qa_direct_db	✅ 7 puta
Trener 2025/26?	rag_qa_direct_db	✅ Victor Sanchez
Proračun PGZ 2026?	rag (llama-4)	✅ 406,9M EUR
Trofeje HNK Rijeka?	rag_qa_direct_db	✅ 2× HNL, 7× Kup, 1× Superkup

📁 KEY FILES MODIFIED

/opt/rinet-gpu/dabi_orchestrator_v3.py — sve fix-ove
/opt/ai-rinet/RINET_HANDOFF_FORENSIC_v3.md — 346 linija forensic
/etc/systemd/system/lora-finetune.timer — DAILY 03:00

🔄 PENDING (sljedeća sesija)

Frontend UI za upload — backend gotov, treba React
GDPR cookie consent banner za production
Server B order — Hetzner AX42-U €72/mj kad PGŽ plati
Verify next LoRA training completes 03:10 sutra
Auto-publish to portal.rinet.one — DABI persona daily articles
Investigate VRAM contention — vLLM disabled, GPU 51% iskorišten

🗝️ OPERATIVNI CHEAT SHEET

# Bridge
curl -sX POST https://api.rinet.one/bridge/exec \
  -H "X-API-KEY: rinet-yS4ZnKlwUqsjk" \
  -d '{"cmd":"..."}'

# DB
PGPASSWORD='R1net2026!SecureDB#v7' psql -U rinet -d rinet_v3 -h 127.0.0.1 -p 6432

# LoRA test
curl -X POST http://localhost:8765/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"Q?"}]}'

# Reranker
curl -X POST http://localhost:8099/rerank \
  -d '{"query":"Q","documents":["d1","d2"],"top_k":3}'

# MCP SSE
curl http://localhost:8810/sse

# Restart orchestrator
systemctl restart dabi-orchestrator-v3 && sleep 25

# LoRA timer status
systemctl list-timers lora-finetune

3.7 KiB Raw Blame History Unescape Escape