Changelog

All notable changes to this project will be documented in this file.

[0.2.1] - 2026-03-18

Heungkuk Life Insurance official CI favicon (frontend/public/favicon.ico)
Korean IME (hangul) composition handling — prevents last character from being left behind on Enter

Smart PDF Parser (server/rag/parsers.py)
- Multi-strategy parsing: text → OCR → table → VLM
- RapidOCR (ONNX) for scanned PDF pages (Korean + English)
- pdfplumber table detection → markdown conversion
- Ollama VLM cloud model for image/chart descriptions
- Per-page metadata: parse_method, parse_quality, has_tables, has_images
OCR/VLM configuration: VLM_MODEL, OCR_ENABLED, VLM_ENABLED environment variables
use_ocr and use_vlm options in ingestion API
poppler-utils system dependency in Docker image
New dependencies: rapidocr-onnxruntime, pdfplumber, pdf2image, Pillow, numpy
Bilingual documentation: README.md (English) + README.ko.md (Korean)
CHANGELOG.md

ingestion.py: Replaced simple pypdf extraction with smart multi-strategy parser
Document metadata now includes parsing method and quality score
Updated README.md with comprehensive project documentation

Initial release: Regulation RAG Multi-Tenant Chatbot
Backend: FastAPI + LlamaIndex + ChromaDB
- CitationQueryEngine for source-referenced answers
- Multi-tenant vector store (per-tenant ChromaDB collections)
- PDF ingestion pipeline: parse → chunk → embed → store
- Tenant CRUD (JSON-based storage)
- Health check with Ollama connectivity status
Frontend: React + Vite + Tailwind CSS
- Chat screen with citation display
- Documents screen with title/summary/page listing
- SPA routing served from FastAPI
Embedding: HuggingFace intfloat/multilingual-e5-small (in-container, no Ollama dependency)
LLM: Ollama integration (cloud model support: nemotron-3-super:cloud)
Docker Compose deployment (Colima compatible)
Korean system prompt for regulation-specific responses