Your agent gives its sharpest answers when each Knowledge Context is structured well. When you have a large content library, it’s tempting to load everything into one context, but the number of files you upload isn’t what shapes answer quality. This guide shows how an agent reads a context, and how to structure one so answers stay accurate.Documentation Index
Fetch the complete documentation index at: https://docs.mindset.ai/llms.txt
Use this file to discover all available pages before exploring further.
How an agent reads a knowledge context
When you add a file to a Knowledge Context, the agent doesn’t read the whole file every time someone asks a question. During upload, Mindset AI breaks each file into small passages of around 250 words, then builds a searchable index of what those passages mean. When a user asks something, the agent searches that index, pulls back the handful of passages closest to the question, and writes its answer from them. Each sentence links back to its source. Two things follow from this. The agent never loads your whole library at once, so a large context doesn’t fill its working memory. And the real unit of knowledge is the passage, not the file.What a passage looks like
Say you upload a 40-page employee handbook. One passage pulled from it might read:Full-time employees receive 25 days of paid annual leave per year, accrued monthly. Leave requests should be submitted at least two weeks in advance through the HR portal. Up to five unused days can be carried into the following year, and anything beyond that is forfeited at year end.When someone asks “how much holiday do I get, and can I carry it over?”, the agent finds this one passage and answers from it. It doesn’t read the other 39 pages. That focus is what keeps the answer both fast and accurate.
Why the number of files is the wrong measure
Files vary widely in how much searchable content they hold, so a file count tells you little about the true size of a context. A one-page FAQ might produce a single passage. A two-hour webinar transcript might produce a few hundred. A dense technical manual holds far more retrievable detail than a slide deck with ten words per slide. Here’s a rough illustration of how different content breaks down:| Content type | Rough passages per file | Why |
|---|---|---|
| Short FAQ or memo | 1 to 3 | Little text, quickly covered |
| Standard report (10 to 20 pages) | 20 to 60 | Moderate density |
| Long manual or book (100+ pages) | Several hundred | Dense and detailed |
| Webinar or meeting transcript (1 to 2 hours) | 100 to 300+ | Speech produces a lot of words |
| Slide deck | A few, often sparse | Relies on visuals, little text |
| Spreadsheet or scanned document | Varies widely | Depends how cleanly the text extracts |
These ranges are illustrative, not exact. The point is the spread: two contexts with the same file count can differ a hundredfold in real content. Plan a context around passages and information density, not file count.
A large context stays fast, but accuracy can slip
Because the agent searches an index instead of reading files one by one, search speed barely changes as a context grows. A context holding 100,000 passages returns results about as fast as one holding 1,000. Accuracy is what changes. As more passages pack into one context, many start to look alike in meaning, and the search has a harder time telling the right passage from one that’s merely related. Answers can quietly lose precision even when nothing looks broken and nothing slows down. Mindset AI runs extra quality filtering after the search, but it can only work with passages the search already found, so it can’t recover a strong passage that got crowded out.How big should a context be?
There’s no single magic number, because it depends on information density and how varied your content is. As a working guide:- A focused context covering a few thousand typical documents stays sharp.
- Treat around 10,000 files as a soft ceiling for typical business content.
- Above that, especially with mixed-topic content, split into several focused contexts rather than one large one.
Split a large library into focused contexts
An agent can use more than one Knowledge Context at once, and it searches across all of them when it answers. Splitting your library costs you nothing in coverage and gains you accuracy. Group content the way your users ask about it, for example by product line, department, region, or document type. Each context stays smaller and sharper, so each search runs over cleaner, more relevant material.Before you load a large library
- Plan around passages and density, not file count.
- Group content into focused contexts rather than one large one.
- Scope each agent or session to only the contexts it needs.
- Load a representative sample first and test answers against real questions.
- Re-check answer quality after any large content addition.