Working with LLMs from Zero
by Sergey Okinchuk

Chapter Four · Your Own Material

Giving It Your Documents

The model only knows what is in front of it, and until now that has been whatever Tessa could type. This chapter puts Waymark's own documents in the window — a contract too big for it, a season of reviews, a table of numbers — and teaches the craft of choosing what goes in, checking what arrived, and knowing which jobs belong to a different tool entirely.

Everything so far has been Tessa and a chat box: asking well, and knowing when not to believe the answer. This chapter brings in the filing cabinet. Waymark's real material — a 120-page supplier contract, a season of guest reviews, a bookings table, a policy folder nobody has read end to end — goes in front of the model, and the work changes character. Answers stop coming from what the model learned in general and start coming from what Waymark actually wrote.

Five pages, one skill underneath. Getting material in, and knowing what arrives (pasting, attaching, and the extraction step that can quietly lose half a document). Fighting a 120-page contract into a window that will not hold it, by finding the three pages that answer the question. The three verbs that do most of an office's work — summarize, extract, classify — tried on 2,300 reviews. Numbers, where fluent text and correct arithmetic part company, and where a spreadsheet takes the job back. And then the chapter's quiet punchline: the routine Tessa built by hand — find the right text, put it in the window, ask — is the same pattern engineers build into products that answer from thousands of documents. She learns the name for something she can already do.

The collision this chapter is about
120 pages against one finite window
What Tessa has on her desk120 dense pages · about 90,000 wordsroughly 120,000 tokenswill not go in whole
What goes in the windowthe section on cancellationsthe cross-reference in the definitionsthe question, asked against those
The skill was never getting everything in front of the model. It is finding the part that answers the question — and then asking the model to quote the clause it used.

Topics in This Chapter

Topic 16
Paste or Attach
Two doors, one room: whatever you send arrives as tokens. What the extraction step does to a PDF, why a scanned page arrives empty without complaint, and the ten-second spot-check that proves what actually made it into the window.
Concept
Topic 17
The 120-Page Contract vs the Window
90,000 words, roughly 120,000 tokens, and one question about group cancellations. Napkin token maths, cutting to the sections that matter, working through it in chunks, and why answering from a summary loses the clause you needed.
Concept
Topic 18
Summarize, Extract, Classify
The three verbs behind most office work with a model, run on 2,300 guest reviews. Summaries with an audience, extraction that generates rather than copies, buckets you define instead of themes it invents, and a trial batch of fifty.
Concept
Topic 19
Tables and Numbers
A confident ranking with wrong arithmetic in it. Why a text predictor produces plausible digits, the request that improves the sums and lets you check them, the echo-back check for pasted tables, and where the spreadsheet takes over.
Concept
Topic 20
Search Plus Context: the Intuition
Find the right text, put it in the window, ask — the habit this chapter built, and the pattern behind every tool that answers from your own manuals. What a search step in front of the model buys, and how a wrong fetch poisons a fluent answer.
Concept