Why context matters in chat moderation

One sentence, two meanings

Previous messages	Message	Without context	With context
“anyone seen the finale?” – “not yet, tonight!”	“lol he dies at the end”	harmless	spoiler
“my goldfish is old, I'm worried”	“I think he dies at the end, sorry”	spoiler?	harmless
“you lost again 😂” – “rematch?”	“you're going down”	threat?	friendly banter
“stop writing to me” – “I said stop”	“I know where you live”	ambiguous	threat

The message alone does not decide. The conversation does.

What context fixes

• False alarms on jokes between friends, gaming banter, replies that quote someone.
• Rules that depend on the topic: spoilers, off-topic messages, selling in the wrong channel.
• Escalation: a message that is mild alone but is the third unwanted contact in a row.

What context must not do

The score is about the last message. If someone else insulted you two messages ago, your polite reply must not be flagged as an insult. A good model uses context to understand the message, not to transfer the previous messages' harm onto it.

How much context?

In chat, meaning usually depends on the last few messages. Sending 5–10 previous messages, oldest first, covers almost every case while keeping requests fast. Very long histories cost more and rarely change the verdict.

Why it is hard for models

Context-aware training data is rare: most public datasets contain single messages. Models must see many examples where the conversation changes the right answer, in both directions, to learn to use it. That is one of the main goals of the next VigiMod version.

In VigiMod

Send the previous messages in context, oldest first, up to 10. The scores remain about text; the context only helps to read it. See the documentation and the Discord guide for an example that fetches the last messages of the channel.