---
title: "Why context is the product — Context Field Notes"
description: "Models are interchangeable. What you feed them is not."
canonical: "https://press.getalchemystai.com/books/context-field-notes/why-context"
last_updated: "2018-10-20"
book: "Context Field Notes"
---

# Why context is the product

Every team that ships an assistant eventually discovers the same thing. The model is rarely the bottleneck. The bottleneck is what the model is allowed to see when it answers. A strong model with the wrong context gives a confident, wrong answer. A modest model with the right context usually gives a useful one.

## The three questions

Before any architecture diagram, it helps to ask three plain questions about a request. What does the system need to know to answer it well? Where does that knowledge live today? How stale is it allowed to be? Most context problems are an honest answer to one of these that nobody wrote down.

The first question is about scope. A support assistant answering a billing question needs the customer's plan, recent invoices and perhaps the last two tickets. It does not need the full product manual. Sending more than is needed is not free: it dilutes attention, costs tokens, and makes failures harder to explain.

The second question is about ownership. Knowledge that lives in a CRM, a wiki and somebody's inbox will drift apart. The assistant inherits that drift. If two sources disagree, the model will pick one, and you will not like how it picks.

The third question is about time. Some facts are stable for years. Others are wrong an hour after they were true. Treating both the same way is the most common source of answers that were correct last week.

## What changes when you take it seriously

Teams that treat context as a product start versioning it. They keep a record of what was retrieved for each answer, so that a bad answer can be replayed and explained. They write tests that assert on retrieved context, not just on final text. And they budget context the way they budget latency: deliberately, per request type.

> The cheapest way to improve an assistant is usually to remove something from its prompt.

None of this is exotic engineering. It is the same discipline that made search and recommendation systems reliable, applied to a new kind of consumer.

---

- Contents: [Context Field Notes](https://press.getalchemystai.com/books/context-field-notes.md)
- Next: [Layers of memory](https://press.getalchemystai.com/books/context-field-notes/memory-layers.md)
