Tune Presidio: strict mode masks names+orgs for NDA, light mode skips them for research

Swap pdfplumber → pypdf for upload parsing (3x faster text extraction)

Handle null content in LLM response (NoneType len error)

Auto-fallback to mercury when minimax fails (chat + stream)

Skip PERSON/LOCATION/ORG in PII masking — only mask contact & financial IDs

Safe minimax defaults: no thinking, temp=0.3, max_tokens=4096

Profile-aware API key: mercury uses INCEPTION_API_KEY, minimax uses LLM_API_KEY

Minimax uses LLM_BASE_URL env var; mercury always uses inceptionlabs.ai

Remove LLM_BASE_URL env override — use profile's own base_url

Keep only mercury + minimax profiles; remove custom profile logic entirely

Drop custom profile as default — always default to mercury; minimax remains selectable

Speed up analyze: use black_box mode (no thinking, 2K tokens) instead of white_box

Simplify document_agent: drop JSON parsing, return raw LLM output; fix llm.py syntax errors

feat: query-driven analyze + context window bump

fix: thinking.type enabled for analyze mode + status event

fix: increase httpx timeout 120s -> 300s/600s for thinking models

fix: thinking.type enabled instead of adaptive for NVIDIA API

feat: minimax temp=0.1 + thinking adaptive + json output for analyze

fix: handle LLM_BASE_URL with or without /chat/completions suffix

feat: provider-based LLM profiles + per-request model switching

feat: add page_width/page_height to citations for frontend bbox scaling + combined bbox+dimension extraction

fix: use pdfplumber text for chunking to ensure accurate page numbers for citations

fix: tune Mercury 2 params — use valid temp range (0.5-0.75), instant reasoning for black-box, high for white-box, reasoning_summary=false for factual paths

fix: keep MarkItDown for chunking, pass pdfplumber page texts separately so page numbers align for bbox extraction

feat: per-mode reasoning_effort (high for white_box/analyze) + max_tokens 5120

fix: handle None text in guardrails check_input (crashed on analyze with null question)

feat: on-demand pdfplumber bbox extraction for citation highlighting

feat: harden prompt injection guardrails with 11 detection layers + fix missing guardrails on /analyze

perf: upgrade embedding to all-mpnet-base-v2, add cross-encoder reranker, LRU cache, enable PII by default

production hardening: circuit breakers, graceful degradation, health checks, parallel uploads, HF Spaces deploy config

feat: /analyze now fairly samples all selected documents, adds cross_document_comparison + per_document_summary fields to response

fix: enable NeMo Guardrails with correct parameters.api_key/parameters.base_url, add PII detection to guardrails flow, set PII_ENABLED=true

fix: correct Inception API base URL (inceptionlabs.ai, not inception.ai)

fix: add generic LLM_API_KEY/LLM_BASE_URL/LLM_MODEL for any OpenAI-compatible API; fall back from broken api.inception.ai

fix: handle LLM connection errors gracefully (SSL/connect errors → 502); convert Analyze 500 → 502 with clear message

fix: 15 critical/high bugs across codebase

fix models.py: match Neon schema varchar sizes, add email unique constraint