v2.12.0 - PDF Ingestion, AI OCR Fallback & LiteLLM Model Discovery

What's New in v2.12.0

📄 PDF Extraction with PyMuPDF & Multimodal AI OCR Fallback

• High-Performance Digital Extraction: Fast text and layout extraction via PyMuPDF with per-page metadata, table structure preservation, and header/footer normalization.
• Vision AI OCR Fallback: Automatically invokes upstream multimodal models (e.g. Gemini 2.5 Flash, Qwen3-VL, Claude 3.5 Sonnet) for scanned, image-only, or dense technical PDFs (such as ASD-STE100 specifications) with page-level rendering.
• PDF Preview & Inspection Modal: Interactive UI modal in Local Storage allowing operators to inspect page text, review AI OCR extraction quality, view generated semantic chunks, and confirm ingestion before writing vectors to Qdrant/pgvector/Chroma.
• MCP Storage Tool Integration: Extended manage_local_file MCP tool with preview action and formatted multi-page PDF reading.

🧠 Dynamic LiteLLM Model Discovery & Hot-Reloadable AI Configuration

• Live Model Discovery: REST endpoint GET /admin/api/models/discover queries upstream LiteLLM gateways (GET /v1/models) and categorizes models into Dense Embeddings, Vision OCR, and General Chat models.
• UI Model Selectors & Discovery Controls: Replaced static text inputs with dynamic capability dropdowns and custom manual write-in options in the Settings tab.
• Relational Metadata Persistence: Model selections (dense_model, sparse_model, vision_ocr_model, chat_model, litellm_url, litellm_api_key) persist to SQLite/PostgreSQL system_metadata and hot-reload in-process without container restarts.

🩺 Vector Database Health Check Fix

• Directly interrogates the active vector store's .health_check() method instead of relying solely on connection status flags, correctly reflecting operational status in the dashboard.

🧪 Automated Test Verification

• Pytest Backend: 492 passed (100%).
• Vitest Frontend: 286 passed across 27 test files (100%).
• Automated CI/CD: End-to-end GitHub Actions workflow validation.