Add MiuCare identity reply + HF Spaces deploy guide

Add HF Spaces Docker deploy

docs: add "refresh to newer data" steps for collaborators

docs: simplify README + add FE multi-turn integration guide

docs: refresh README + HF dataset card to current corpus (76,487 chunks, 11 sources)

tune: confirm retrieval history window = 2 turns (FE-managed multi-turn)

feat: history-aware retrieval for multi-turn coreference

fix(classifier): decouple severity from health-only topic; petchoice crawl

crawler: add verified petchoice.vn source for breed body-condition gap

eval: faithfulness spot-check on health subset — no regression (5.00/5)

perf(retriever): cut rerank latency ~80s->~28s (candidates 20->10, maxlen 512->384)

add scripts/test_api.py — CLI client to test /chat without a frontend

fix(classifier): bound 'ho'/'sán' substrings — recover behavior coverage

docs: update eval section with v7 results (citations 45/45, ~5s latency)

perf(llm): disable gemini-2.5 thinking — fixes truncation + ~10x latency

fix(llm): raise max_output_tokens 1024 -> 2048 to stop truncating cited answers

fix(safety): make _compute_needs_vet query required; pass query in eval harness

needs_vet: harden intent-gate with 20-case emergency stress set

needs_vet: add query intent-gate to cut emergency-banner over-trigger 8→1

fix(prompts): stop false give-ups + enforce citations in list answers

fix(classifier): replace noisy bare "cá" keyword causing topic mislabels

chore: regenerate uv.lock after google-genai migration

docs: record topic-routing voting investigation (no-op)

feat(llm): migrate to google-genai SDK + quota resilience

docs(safety): document why classifier-level severity fix is rejected

docs(eval): record full judged eval — faithfulness 4.86, helpfulness 4.69

test(eval): score the shipped safety behavior (rerank gate + server banner)

fix(eval): make LLM judge robust to 2.5-flash thinking-token truncation

fix(safety): gate needs_vet on rerank relevance + server-side vet banner

fix(llm): split comma/space-separated keys in GEMINI_API_KEY too

feat(llm): rotate across multiple Gemini API keys to multiply free-tier quota

docs(eval): record measured retrieval+safety metrics from external suite

fix(eval): use live Gemini models in judge fallback chain

fix(llm): drop retired gemini-1.5-flash, add gemini-2.0-flash fallbacks

feat(rag): add cross-encoder reranker, enforce citations, external-dataset eval harness

feat: harden safety flag, add API auth + rate limit, relevance gate, tests

docs: add Troubleshooting section with 4 real-world install gotchas

fix(retriever): use transformers AutoModel directly, bypass sentence_transformers

docs(README): note huggingface_hub commit-step hang for large files

fix: HF username is Monmoonluna (capital M), not monmoonluna

feat: HuggingFace Datasets upload/download scripts

docs: comprehensive README update — final state of v1

feat: Windows Task Scheduler-based ingest watchdog

fix(finish_pipeline): strip Unicode em-dash that broke PowerShell parser

fix(finish_pipeline): drop Python sub-spawn for ChromaDB count

feat: ingest auto-restart wrapper + eval chain

feat: cache-hit short-circuit + auto-restart wrapper + omnibus pipeline

feat: 5 behavior-focused VN sources for phase 2 (3,383 new URLs)

fix(classifier): require ≥2 hits to flip topic + ingest metadata-only mode

test: eval script + 30-query bug-finding suite