Enterprise RAG has long been sold as a ready‑made pipeline: chunk, vector store, top‑k, optional rerank, LLM. In practice, many deployments fall short, producing vague citations and unreliable answers. The article argues the real win comes from deep engineering, domain knowledge, and a clear understanding of the documents themselves.

The author shows a 100‑line script that parses PDFs, retrieves the top three pages by cosine similarity, and feeds them to an LLM with a Pydantic schema. The script outputs structured answers with line‑level citations, outperforming complex, multi‑layered systems that stack rerankers and query‑rewriters.

By separating extraction from generation, the pipeline preserves an audit trail and keeps the LLM focused on validating retrieved facts. This approach aligns with the original 2020 RAG concept, which emphasized grounding answers in fresh corpus data rather than relying on pre‑trained memory. For companies with a handful of core documents and recurring expert queries, such a lean architecture delivers transparency and reliability.

The series proposes a four‑brick pipeline: parsing, question parsing, retrieval, generation, optionally annotating PDFs. It offers a concrete, verifiable baseline that demonstrates where production systems often over‑engineer and lose sight of the core problem.