Configuring Advanced Context

Set How Much Source Material Reaches the Model

Select Agents in the left-hand navigation, open your Agent, then choose Knowledge from the row of pages across the top and select the Advanced tab.

Three controls decide what the Agent has to work with:

• Max Sources caps how many source passages are included, from 0 to 25. More passages give the Agent more to draw on and make each prompt more expensive. Setting it to 0 stops source material reaching the Agent at all.
• Relevance Threshold is the minimum relevance score a passage must reach to be included, from 0 to 1 and displayed as a percentage. Raise it when the Agent cites material that is only loosely related. Lower it when the Agent says it cannot find anything on topics you know are covered.
• Source Handling Instructions tells the Agent what to do with those passages. The {passages} merge tag is where they are inserted, so the tag must be present for the Agent to see them at all.

The two numbers work against each other. A high Max Sources with a high Relevance Threshold often yields fewer passages than the cap suggests, because the threshold filters them out first.

Decide Whether the Agent May Search the Web

Choose the Behavior page and select the Abilities tab.

Enable Live Web Search lets the Agent look up current information when its own sources do not cover a question. Turning it on unlocks three further controls, all of which stay greyed out while it is off:

• Include Domains restricts searching to the domains you list. Leave it empty to allow all domains.
• Exclude Domains blocks specific domains. Use it to keep the Agent away from sites that conflict with your positions.
• Web Search Tool Instructions tells the Agent when to reach for search rather than answering from what it already has.

Turning the toggle on fills the instructions field with a working default, so you get sensible behavior before you change anything. Turning it back off clears that text again, but only if you have not edited it.

Set How Much Conversation the Agent Remembers

Still on the Behavior page, select the Memory tab.

• Max Memories is how many previous prompt and response exchanges are passed back to the model, from 0 to 25. This is what lets someone ask a follow-up question without repeating themselves. Setting it to 0 makes every prompt independent, which suits a single-question kiosk and frustrates anyone holding a conversation.
• Persist Conversation Sources keeps the source passages found earlier in a conversation available for later turns. Without it, each turn searches afresh, and a follow-up question phrased vaguely may retrieve nothing useful.

Adjust the Language Handling Instructions

Choose the Context page and select the Language tab.

Language Handling Instructions tells the Agent how to behave with respect to language, and it supports the {language} merge tag, which resolves to the language the user is working in.

This field is prefilled with a working default. Edit it only when you need behavior the default does not give you, such as instructing the Agent to keep certain proper nouns untranslated.

This is a different setting from the supported and default languages covered in Configuring Languages. Those decide which languages the Agent offers. This decides what the Agent is told about the language once one is chosen.

Adjust the Scripture Handling Instructions

Still on the Context page, select the Scripture tab.

Scripture Handling Instructions governs how the Agent quotes and references scripture, and it supports the {bible} merge tag, which resolves to the translation in use for the request.

This is where you state house rules on scripture, such as how many verses a response should cite or whether the Agent should give the reference before the quotation. Which translations are available is set separately, in Configuring Bible Translations.

Give the Agent Situational Awareness

Still on the Context page, select the Miscellaneous tab.

Miscellaneous Context Instructions is where date and location awareness comes from, and it carries the widest set of merge tags of any field on these pages.

Unlike the Language and Scripture fields, this one starts empty, and an empty field means the Agent receives none of this context at all. That is why an Agent with nothing here will confidently state the wrong date if asked: with no {date} tag reaching it, the only date it has is the one implied by its training data.

Write plain instructions and place the tags inside them, for example:

## CURRENT CONTEXT:
- Today's date is {date}.
- The user is in {city}, {region}, {country}.
- Use this only when the question depends on the current date or the user's location.

The last line matters. Supplying a tag makes the information available; it does not tell the Agent when the information is relevant.

Location tags resolve to an empty string when the user’s location cannot be determined, which leaves a sentence such as “The user is in , , .” in the prompt. Phrase location instructions so they still read sensibly when the values are missing, or keep location on its own line rather than mixed into a sentence you rely on.

Save And Verify Against Real Questions

Select Save Changes on each page you changed. Each of the three pages saves independently, so a change on Knowledge is not saved by pressing Save on Context.

Then open the Agent’s live URL and test the settings you touched:

• Ask a question whose answer depends on today’s date, to confirm the {date} tag is reaching the Agent.
• Ask a question, then a vague follow-up such as “why?”, to confirm Max Memories is carrying the conversation.
• Ask a question you know your sources cover, and use View Sources on the response to confirm passages are being retrieved under your current threshold.