At the moment if the set of messages for inference has more than one system message, only the last system message will be put into the Anthropic API system parameter.
During some testing of group chats I noticed there are multiple system messages for speaker selection in auto mode. So this meant the speaker selection was losing context and failing.
This PR simply concatenates the system messages together and it proves effective.
Why are these changes needed?
Group Chat doesn't work effectively with Anthropic models when in auto mode without it.
At the moment if the set of messages for inference has more than one system message, only the last system message will be put into the Anthropic API
system
parameter.During some testing of group chats I noticed there are multiple system messages for speaker selection in auto mode. So this meant the speaker selection was losing context and failing.
This PR simply concatenates the system messages together and it proves effective.
Why are these changes needed?
Group Chat doesn't work effectively with Anthropic models when in auto mode without it.
Related issue number
None raised.
Checks
(3392)