fix: correct Llama 3.3 70B model ID (llama3.3-70b → llama-3.3-70b)

feat: migrate LLM backend from HF InferenceAPI to Cerebras

fix: switch to non-streaming LLM call for mobile stability

feat: mobile-responsive CSS — stack chat+sidebar vertically on small screens

perf: reduce CE_POOL 12→6 for faster reranking

fix: append Traditional Chinese reminder to user_msg when Chinese detected

fix: streaming error handling + strengthen Traditional Chinese instruction

fix: revert reranker to MiniLM-L-6-v2 (stable)

perf: swap reranker to BAAI/bge-reranker-base (278M, Chinese+English)

fix: enforce Traditional Chinese responses + downgrade reranker to base

feat: Traditional Chinese UI + upgrade reranker to mxbai-rerank-large-v2

feat: multi-LLM support with A/B testing dropdown (4 models)

feat: cross-encoder reranking (ms-marco-MiniLM-L-6-v2) — 3-stage pipeline: BM25+FAISS → RRF → CE rerank

feat: hybrid search — BM25 + FAISS via Reciprocal Rank Fusion (RRF)

fix: use gr.State for sidebar input (gr.Markdown is output-only in Gradio 5)

feat: two-column UI — sources sidebar with confidence scores, preview, colour indicators

fix: calibrate confidence scores (L2→60-100% range) + star rating display

fix: guard empty choices in streaming loop (finish chunk)

feat: streaming response + FAISS confidence scores per source

fix: remove all langchain deps — pure FAISS+sentence-transformers+HF pipeline; add metadata.json

fix: pin langchain>=1.0 and pydantic>=2.7 to resolve pydantic.v1 circular import on Python 3.13

fix: eager-load pipeline at startup + loading status indicator

fix: upgrade to Gradio 5.x for Python 3.13 compatibility (audioop removed)

Initial deployment: UTM LLM Assistant (22 docs + FAISS index, LFS)