chore: remove unintended committed files

fix: handle None article in answer_validator startswith

fix: add try/except with traceback logging to /chat endpoint

restart: switch retrieval back to MiniLM (remove INDEX_SUFFIX)

remove /diag debug endpoint

update frontend references to working Vercel URL

add fallback Vercel URL to CORS

fix: EURLEXEmbedder.encode() must return numpy array, not list

fix: EURLEXEmbedder.encode() must accept normalize_embeddings kwarg that SentenceTransformer uses

add /diag endpoint for debugging

fix: pass HF_TOKEN to onnx model download

fix: use suffix variables in download_index instead of hardcoded filenames

fix: add /diagnose endpoint for EURLEX-BERT debugging

fix: add onnxruntime+transformers deps and fix model download repo_type

fix: also upload build_meta_eurlex.json

fix: download chunks.db in merge job

fix: create embeddings directory before saving

fix: correct ONNX model path after download

fix: download ONNX model from Space repo instead of dataset

fix: build-index workflow – direct download, upload chunks_eurlex.db

feat: improve CELEX citation compliance and fix build_index bugs

fix: strengthen CELEX citation instructions in LLM prompt

phase-3: fix feedback-analysis.yml, final test pass

phase-2: EURLEX-BERT ONNX INT8 quantization, matrix sharding workflow, EURLEXEmbedder

phase-1: CI workflow, ruff linting, keepalive alerting, VITE_API_URL, pre-commit hooks

phase-0: remove dead code (hfspace/, backend/, render.yaml), update checkpoint scripts

phase-0: add HF Spaces sdk:docker frontmatter to README.md

phase-0: Dockerfile, .dockerignore, package __init__, merged requirements, Phoenix Plan

checkpoint: create app/ directory with consolidated backend (9-stage pipeline)

fix: reduce EURLEX-BERT embed batch size from 1024 to 64 to avoid OOM on 7GB runner

fix: split rebuild into 2 jobs (download + build) to bypass 6h timeout

fix: add --from-backup to skip 5h SPARQL crawl by loading chunks from HF backup

fix: EURLEX-BERT uses transformers, not sentence-transformers — implement mean pooling

fix: orphan docstring text from bad merge caused SyntaxError

fix: revert rebuild condition to boolean check, use --field for dispatch

fix: continue-on-error for backup step so rebuild runs regardless

fix: rebuild condition compares with string 'true'

docs: add comprehensive README with architecture, features, API, deployment

fix: remove nodeVersion from vercel.json, hook AutoExpander into main.py

merge: resolve rollback.sh conflict

feat: complete 4-phase improvement strategy

Incorporate local version improvements: FAISS+SQLite storage, optimized build process, automation scripts

feat: optimize build pipeline with evidence-backed improvements

checkpoint: pre-PNP-refactor state

chore: trigger fresh Vercel deploy

fix: hardcode API URL to HF Space

feat: add GitHub Actions keepalive (5min) and index refresh (hourly)

fix: pin huggingface-hub to <1.0 for transformers compat

feat: add HF Space deployment config, update frontend API URL

update render.yaml for deployment