> For the complete documentation index, see [llms.txt](https://docs.xenioo.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.xenioo.com/ai/large-language-models/resources/settings.md).

# Settings

The **Settings** tab, next to the Resources list, controls how your model turns its resources into an index and how it searches that index when answering. Unlike the options in the Add Resource dialog, these settings apply to **every resource of the model at once:** you configure them once, and they shape indexing and retrieval for the whole knowledge base.

<figure><img src="https://2532694803-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FvGYwUmVjfELEiJzfmWLC%2Fuploads%2FfaGX6szgDOHkfT91EiWV%2Fimage.png?alt=media&amp;token=5b99c82f-bbb0-4663-8dba-fe8cb66245ce" alt=""><figcaption></figcaption></figure>

{% hint style="info" %}
The defaults are tuned to work well for most bots. Change them only when you have a specific reason; for example very large documents, or answers that pull in too much (or too little) context.
{% endhint %}

Every change takes effect the next time a resource is indexed. Existing content keeps its current index until it is reindexed, so adjust the values and then reindex the resources you want to rebuild.

### Indexing

These two settings decide how a document is broken into **chunks,** the small pieces that are actually embedded and searched. Getting the chunk shape right is the single biggest lever on answer quality.

#### Chunk Size (in tokens)

The maximum size of each text chunk a document is split into, measured in tokens.

* **Larger chunks** keep more surrounding context together, which helps the model understand a passage but each match brings in more text, so answers can become broader and use more tokens.
* **Smaller chunks** are more precise and targeted, but risk cutting an idea in half across a boundary.

|         | Value    |
| ------- | -------- |
| Minimum | **200**  |
| Maximum | **8000** |
| Default | **8000** |

{% hint style="info" %}
The maximum of 8000 matches the largest input the embedding model accepts. It also lines up with the **No Split** chunking strategy: a document under this size can be kept whole as a single chunk.
{% endhint %}

#### Chunk Overlap (in %)

How much of each chunk is repeated from the end of the previous one, as a percentage of the chunk size. Overlap keeps context from being lost exactly at the boundary between two chunks, so an idea that spans a split is still fully present in at least one chunk.

* **0%:** no repetition; chunks are strictly sequential.
* **Higher overlap:** safer boundaries, at the cost of some duplicated content and a slightly larger index.

|         | Value                       |
| ------- | --------------------------- |
| Minimum | **0%**                      |
| Maximum | **25%** (of the chunk size) |
| Default | **0%**                      |

### Retrieval

These settings control what happens at answer time, when the model searches the index for the chunks most relevant to the user's question.

#### Limit (documents in vector search)

The maximum number of matching chunks retrieved from your resources to build an answer. A higher limit gives the model more material to work with, but too many weakly-related chunks can dilute the answer and cost more tokens.

|         | Value   |
| ------- | ------- |
| Minimum | **100** |
| Default | **500** |

#### Score Threshold (minimum similarity value)

The minimum relevance score a chunk must reach to be used in an answer. Every candidate chunk is scored by how similar it is to the question; anything below this threshold is dropped.

* **Higher threshold** stricter: fewer but more relevant matches, less noise. Set it too high and the model may find nothing to answer with.
* **Lower threshold** more permissive: more matches, including loosely related ones.

|         | Value   |
| ------- | ------- |
| Minimum | **0.3** |
| Maximum | **0.9** |
| Default | **0.3** |

{% hint style="info" %}
**Limit** and **Score Threshold** work together: Limit caps *how many* chunks are considered, Score Threshold decides *how relevant* they must be. If answers miss information that is clearly in your resources, lower the threshold; if they wander off-topic, raise it.
{% endhint %}

### Auto Exclude (Ban)

Automatically keeps low- or high-value pages discovered by a website or Confluence scan out of the index, based on their estimated token count so you don't have to review a large scan page by page.

Turn on **Enable automatic exclusion**, then set the limits:

* **Min tokens:** a discovered page with fewer estimated tokens than this is banned. Useful for near-empty, boilerplate or error pages. `0` means no minimum.
* **Max tokens**: a discovered page with more estimated tokens than this is banned. Useful for huge, unfocused pages. `0` means no maximum.

A page excluded this way is marked **Banned** in the Resources list, with the reason shown on its status.

{% hint style="info" %}
Automatic exclusion applies **only to pages discovered by a scan** never to a file, URL or Confluence root you added yourself. It also never overrides a manual decision: a page you **banned or unbanned** by hand keeps your choice regardless of these rules.
{% endhint %}

### Inherited settings

When a bot is a **child of a Master Bot** and inherits its resources, this whole tab is read-only: the settings are managed on the Master Bot and updated automatically. Use **Detach** to manage them independently after which the child stops receiving resource updates from the Master Bot.

### Saving

Click **Save** to apply your changes. The button stays disabled while the form has invalid values or while another save is in progress. Remember that new values only affect content indexed **after** the save reindex existing resources to rebuild them with the new settings.