Context & Agents · Knowledge Bases
Knowledge base document
See a document exactly as the agents read it, section by section, and choose which sections they can find.
When a document is added to a knowledge base, agents do not read the original file. They read the text Next pulled out of it, split into sections. This page shows you that text, so you can check that the agents are reading what you expect.
Use it to confirm that an important document came through cleanly, to understand why an agent did or did not quote a passage, and to leave out sections that should never be used, such as an out-of-date appendix.
Before you start
- Open a knowledge base and select a document’s name in the Documents table. Only documents whose Ingest has reached completed can be opened. See Knowledge base detail.
- You need permission to change the knowledge base to turn sections on or off. Without it, you can still read everything.
- Use a wide screen. On narrow screens the outline and the section panel are hidden.
Understanding the page
About this document
The top of the page shows:
- the document’s name, its version, a current badge if it is the latest version, and its type
- a counts line, such as 12 section(s) · 10 enriched · 40 chunk(s) at 500/50 tokens – how many sections it has, how many were enriched, how many chunks it was cut into, and the chunk and overlap size in tokens
- what enrichment wrote about it – an extra pass in which Next summarises the document: a summary, Pull this document when (the situations an agent should reach for it), and its topics
A note underneath reminds you: This is the text the agents read — extracted and reconstructed, exactly. Images, charts and layout are not part of it.
A few terms help here:
| Term | What it means |
|---|---|
| Section | A natural part of the document, such as a heading and the text under it. The documents table calls these Units. |
| Chunk | A short piece of text cut from a section, small enough for a search to match precisely. |
| Overlap | The few tokens each chunk shares with the next, so a sentence is never cut in half without context. |
| Token | The unit AI models count text in, roughly three-quarters of a word. |
| Embedding | Storing a chunk’s meaning so it can be found by what it says, not just its exact words. |
Outline
The Outline on the left lists the document’s sections in order, with page numbers where the file format has them. If it has none, it says This format carries no page numbers.
- Select a section to select it in all three panels.
- excluded marks a section that has been left out of searches.
- A note may say that some short sections were too small to summarise. They are still part of the document.
The text, section by section
The middle panel shows each section exactly as stored. Beside each one you see:
- how many tokens it has, such as 320 tk
- how many tokens were embedded, such as 380 embedded
- no summary, if enrichment did not summarise that section
The selected section
The panel on the right describes the section you have selected.
| Item | What it shows |
|---|---|
| Section | The section’s type and its level in the document’s structure. |
| Summary | What enrichment wrote about the section. |
| Retrieval matches on | What a search will match this section against. |
| Cost of this section | Tokens in the text (the section counted once) and Tokens embedded across its chunks. |
| Used in retrieval | Whether searches can find this section. |
Both cost numbers are correct. Chunks overlap, so a section split across several chunks has its overlap embedded twice, which makes the second number larger.
Leave a section out of searches
Use this for content agents should never quote, such as superseded guidance or a legal disclaimer.
-
Select the section in the Outline or the text.
-
Untick Used in retrieval in the right-hand panel.
A message confirms the change. The section is marked excluded in the outline.
The section stays in the document, and nothing is processed again. To bring it back, tick Used in retrieval again.
The box is greyed out if the section was never enriched, or if you cannot edit this knowledge base.
What to expect
- Excluding a section affects every search against this knowledge base, by every agent that can read it.
- If the document cannot be loaded, the page says This document could not be read. Select Refresh to try again.
Good to know
- What you see is what agents get. If text is missing here, for example because it was inside an image, the agents cannot read it either. Consider uploading a text version of that content.
- To correct the document itself, remove it from the knowledge base and upload the new copy. See Knowledge base detail.
- Take the tour. The help (?) menu in the top bar replays this page’s guided tour.