Ingestion & OCR Workers
Documents enter the Go API Gateway and are queued via RabbitMQ. Distributed Python workers run OCR extraction, table segmentation, and text tokenization asynchronously.
Chunking & Milvus Indexing
Extracted text is partitioned into semantic chunks with overlapping boundaries, embedded via all-MiniLM-L6-v2, and indexed into a local Milvus HNSW vector database.
Grounded RAG & Citation
User queries retrieve the top-k most relevant document chunks. The answer synthesizer verifies exact clause references and returns direct page links with zero fabrication.