Building AI That Knows Your Data

On a Tuesday morning in March, the chief executive of Northwind Industries asked a simple question: have we ever agreed to a liability cap below one million dollars? The answer existed. It was written down, signed, and filed. Finding it took three days, and not finding it in time cost forty thousand dollars.

Every organization has a version of that Tuesday. The knowledge is real, it survived, and it is spread across a million files in a hundred formats, organized by whoever was closest to the filing cabinet that day. An organization knows more than anyone in it. That is the problem this book exists to solve.

Building AI That Knows Your Data teaches you to build Retrieval-Augmented Generation the way it should be learned: from scratch, from the documents up, with no magic imports doing the thinking for you. You will build a working system for Northwind, a fictional 240-person manufacturer with thirty years of documents and no way to search them, alongside a cast you will come to know by instinct: the CEO with the only acceptance test that matters, the operations manager who has been the company's search engine for twenty years, the IT lead who builds the machine, and the in-house lawyer who trusts nothing until it shows its sources, and is right to.

Volume I builds the offline half: everything that happens to documents before anyone asks a question. You will crack open PDFs, Word files, spreadsheets, emails, and one scanned contract from 1994 that behaves like a sealed envelope. You will clean what you extracted, cut it into pieces a search engine can love, turn text into searchable meaning, and store a million vectors somewhere sensible. By the final chapter, a question in plain English finds the right passage across formats, vocabularies, and decades, with a score, a label, and a page number.

What you need: a computer that can run Python. No GPU, no cloud account, no budget. Every model in this volume is small, free, and runs locally, which is not a compromise but a lesson: the first serious question in enterprise AI is where your data is allowed to go, and this book answers it, for the pilot, with nowhere.

Every printed line of code was run before it was printed. By the end of this volume, you will have built a complete offline retrieval pipeline: extraction, cleaning, chunking, embedding, and storage, with a working search index at the end of it.

Volume I is the foundation. Volume II opens the front desk, where the system stops finding passages and starts answering questions with citations.