Poor PDF Parsing, Not the LLM, Often Causes RAG Pipeline Failures

A developer stress-testing a retrieval-augmented generation (RAG) pipeline on table-heavy Korean documents found that the PDF parser, not the retriever or language model, was the primary source of incorrect answers. When parsers flatten structured tables into raw text, the relationships between headers, cells, and values are destroyed, causing the LLM to hallucinate connections between unrelated data. Benchmarking revealed that PyPDFLoader corrupted Korean encoding and lost all table structure, while pdfplumber preserved row and column boundaries locally without relying on external APIs. Commercial tools like LlamaParse offered the best extraction accuracy but were ruled out due to the security risk of sending sensitive documents to third-party services. The developer concluded that improving parsing quality — including converting tables to clean Markdown before chunking — delivers greater RAG accuracy gains than tuning embeddings or prompt engineering.