RAG is not the same thing as knowledge

Related build: insurance_nerds

InsuranceNerds is a chat interface for insurance CSRs. They ask questions, the system finds relevant chunks from SharePoint docs, GPT-4o synthesizes an answer. Classic RAG.

PLG Knowledge ingests the same kind of source material — PDFs, Word docs, Excel files — but instead of embedding chunks, it uses structured extraction to build a relational database: deals, people, organizations, provenance for every field, a reconciliation layer for when documents conflict.

I built both of them and called them both "RAG" for longer than I should have. They're not the same thing.

What RAG actually does

RAG finds relevant text and synthesizes it. It's fast to build, flexible, and good at answering open-ended questions where the answer is somewhere in the documents.

It struggles when you need precision. "What's the close rate for brokers we've worked with more than twice?" is not a question vector search handles well. It's a SQL question. The answer lives in a structured record, not in a passage of text.

What structured extraction does

The PLG system prompts an LLM to read each document and produce a typed, validated schema. Every deal has fields. Every field has a source document and page. When two documents disagree on a deal value, the reconciliation layer picks the more authoritative one and notes the conflict.

The result isn't a search index. It's a database that didn't exist before the documents were processed. You can query it like a database.

The tradeoff

RAG is easier to build by an order of magnitude. The PLG system took five active development days just to get the extraction pipeline stable. Handling scanned PDFs with Tesseract, dealing with structured Excel files differently than prose PDFs, writing evals against golden sets to verify the extraction is accurate — none of that is trivial.

But for the right use case — where the user needs specific, reliable, queryable facts, not just relevant passages — RAG is the wrong tool. Knowing which one to reach for before you start building is worth a lot.