> ## Documentation Index
> Fetch the complete documentation index at: https://docs.clarky.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Context Settings

> Tune retrieval, watch the voice budget, and test what your agent would use

## Retrieval

Documents that are not pinned are split into passages. For every incoming message the closest passages are added to the prompt.

* **Relevance threshold** (0 to 100%): how close a passage has to be to count. Lower finds more, higher is stricter. Around 30% works well for most sites.
* **Passages per message** (1 to 20): how many passages can be added. More gives more coverage and costs a little more per reply.

## Voice budget

Phone calls have no per-message retrieval. Everything checked for **Voice** is sent to the voice provider in full when a call starts, so on voice every document behaves like a pinned one. The budget (about 40,000 characters) keeps that prompt small, mostly for latency: a bigger prompt means a longer pause before each spoken reply, and a higher cost per minute.

The meter on this page (and on the Documents screen) shows how much of the budget your voice documents use. If you go over, the oldest documents are cut off on calls. Website pages never count; they are not used on voice.

## Test retrieval

Type a question a customer might ask and pick a channel. Clarky shows:

* **Always included**: the pinned documents that go with every message on that channel
* **Retrieved for this question**: the documents whose passages cleared the threshold, with a match score and the matched passage

If nothing is retrieved, lower the threshold or add a document that covers the topic. On a real conversation the same question would land in [Q\&A](/features/context/qa) as unanswered.