fix: MultilingualEmbedder respects HF_HUB_OFFLINE env var instead of hardcoded local_files_only=True; start.sh uses set -e for error visibility

Fix: role-type skill detection, OpenRouter base_url support, Dockerfile for HF Spaces with 500-profile sample indexes

feat: align candidate relevance ranking & implement robust Turbo Mode fallback

Fix remote candidate search returning 0 results by mapping preferred_work_mode

fix: app.py scoping error + new Sage+Terracotta theme

feat: complete UI redesign with pastel purple/orange theme

refactor: extract shared query parser module, clean up imports

fix: reranker degradation, skip_reranker bug, eval latency, branch cleanup

feat: full 100k index pipeline - switch from samples to full offset index

UI: zero gradients, peaceful warm palette, borders for elevation

UI redesign: pastel theme, premium card design, progressive loading

feat: final smoke-test fixes + spectacular README + UI bugfixes

style(style): resolve ruff formatting check warnings

perf(agents): parallelize candidate profile loading

perf(core): reuse open file handle in ProfileStore for fast candidate fetches

perf(search): use np.argpartition for fast top-k selection in BM25

perf(search): upgrade FAISS index to IndexHNSWFlat

style(search): format reranker.py to fix E501 line length issue

style(language): format translator.py to fix E501 line length issues

feat(ui): wire search filters from Gradio UI to the orchestrator

feat(language): add HingBERT token classification pipeline for Hinglish NER

feat(language): add MBart-50 translation fallback model

fix(search): implement eager non-blocking reranker model loading and online download fallback

fix(api): wire slider_weights, filters, max_results and remove dead code

fix(core): add default values to MatchScores and slider_weights to SearchRequest

fix(matching): map cultural_fit to semantic_similarity

fix(ranking): remove sync rank and sync _judge_group wrapper

fix(agents): instantiate experience/location models in planner fallback

fix(search): change FAISS index type annotation to faiss.Index

fix(core): wrap OpenAI api_key in SecretStr

fix(agents): compile graph on init and add state filter wiring

chore: fix all ruff lint errors (174 -> 0), clean up unused imports/vars, fix line lengths

Refine skill matching and add original query pipeline

Multi-dimensional scoring upgrade: behavioral signals, career trajectory, skill proficiency, honeypot detection

Cross-encoder enabled, rank.py entry point, multi-query pipeline

Fix e2e test failures: planner/reflector timeouts, bias-detector, listwise-rank API

Fix cross-encoder loading, add offline mode, improve simple query detection

fix: add sys.path for direct execution, fix Gradio 6.0 API change

fix: improve UI robustness and test quality

fix: resolve 20 high/medium bugs across search, agents, core, and extraction modules

fix: resolve critical bugs - evaluation import cycle, bias detector, slider weights, and 7 more high-severity issues

Overhaul: Plackett-Luce ranking, PII anonymization, live fairness dashboard + LLM-free fallback agents

chore: add hackathon submission assets, move slider defaults to YAML config

feat(language): add Hinglish code-mixed processor

feat(agents): turbo mode for short queries, stop words, rationale generator wiring

fix(search): cross-encoder sigmoid normalization, hybrid search cache, exact location filters

feat(extraction): add FieldExtractor pipeline, ProfileStore with lazy loading, and skill matching

feat(perf): switch to Ollama qwen2.5:7b, pre-load ML models at startup

chore: commit remaining changes from previous sessions (PRD, plans, tests)

feat(scripts): rewrite build_indexes and evaluate with streaming and demo mode