Hybrid RAG: BM25+Dense (sqlite-vec/BGE-M3) + cross-encoder reranker (bge-reranker-v2-m3)

Hybrid RAG: BM25+Dense (sqlite-vec/BGE-M3) + cross-encoder reranker (bge-reranker-v2-m3)

Hybrid RAG: BM25+Dense (sqlite-vec/BGE-M3) + cross-encoder reranker (bge-reranker-v2-m3)

Hybrid RAG: BM25+Dense (sqlite-vec/BGE-M3) + cross-encoder reranker (bge-reranker-v2-m3)

Hybrid RAG: BM25+Dense (sqlite-vec/BGE-M3) + cross-encoder reranker (bge-reranker-v2-m3)

Hybrid RAG: BM25+Dense (sqlite-vec/BGE-M3) + cross-encoder reranker (bge-reranker-v2-m3)

Hybrid RAG: BM25+Dense (sqlite-vec/BGE-M3) + cross-encoder reranker (bge-reranker-v2-m3)

Hybrid RAG: BM25+Dense (sqlite-vec/BGE-M3) + cross-encoder reranker (bge-reranker-v2-m3)

Retriever: RRF로 키워드+원본 질문 결합 — LLM이 핵심 주제어 누락(e.g. '처방전 보관기간' → ['보관기간']) 해도 원본 질문이 안전망. cases·guides 동일 적용

LLM router prompt: 핵심 주제어 보존 규칙 추가 — '처방전 보관기간' 같은 짧은 질문에서 주제어(처방전) 누락 시 CCTV 등 다른 도메인이 hit되는 문제 수정

feat: switch to Unsloth UD Q2/Q3/Q4 lineup; drop Qwen presets

feat: drop bartowski Q4_K_M Gemma preset (duplicate of unsloth on ZeroGPU)

feat: swap bartowski Q4_0 → unsloth UD-Q4_K_XL, set as default

ui: split-view tab title → '개인정보 상담 AI'

feat: drop unused presets (qwen2.5-1.5b, qwen3-1.7b)

ui: hide '초기화' button from references panel header

ui: sort cards with explicit (근거N) chip to the top of references panel

ui: simplify references panel — drop 'adopted' tier, keep 'LLM 전달' + (근거N) chip

ui: drop model selector from split-view references panel (redundant w/ Open WebUI dropdown)

feat: add Gemma 4 E2B Q4_0 preset (Metal-friendly 4-bit variant)

feat: model preset selector (Gemma 4 + 4 Qwen variants)

Debug: console.log on parent message listener

Listen for kpaa-route postMessage from Open WebUI iframe -> auto clear refs

Auto-clear refs on new chat (no history + non-meta first user message)

Add /api/clear-references + 🔄 초기화 button on right pane

Rename model display: kpaa-privacy-ko -> 개인정보 미니 상담 AI

Use generator pattern (cross-process yield via spaces res_queue) — ZeroGPU forks process so streamer queue isn't shared

Force streamer.end() in finally — transformers 5.x doesn't always auto-call it

Initial backend code: src/kpaa, runtime data, requirements