From posts that I’ve read, the Perchance AI Character Char (ACC) will only load Lore files via URL into the main AI character.
Does this mean that it ignores Lore URLs in other characters’ Lore entry box (won’t load them), or appends those URLs as well into the main character’s Lore?
At first, I thought this was an easy fix: in the main character Lore URL box, just enter the Lorebook URLs of all the characters (e.g., “Mike.txt”, “Tim.txt”, etc.) and they’d all get loaded.
However, my best understanding of how Lore is used is that when a user message is formed and sent to the AI LLM, the ACC will (always? occasionally?) select one (or more?) Lore entries at random to help flavor the AI response.
I think this means that if other characters’ Lore is mixed in to a single “pool” of Lore entries under the main character, then for a Tim character reply, the Lore sent over to the LLM could be random picks of from the pool that contains “Mike” lore and “Tim” lore, etc. In other words, flavoring the “Tim” reply with unrelated Lore facts about Mike.
Could someone kindly let me know if my thinking above is valid? Thanks!
Thanks for the details, that helps a lot.
I understand about poorly-written Lore entries like “he is a snappy dresser” (“he” is vague). My curiosity was more related to Lore such as: “Tim is the sheriff” “Mike is the bad guy”
If the UI is currently constructing a message to the AI and we are discussing Tim, whether the UI could randomly select the “Mike is a bad guy” line from the Lore entries and it would be totally unrelated to the current topic of Tim.
Am I understanding right that Lore isn’t so much “chosen at random”, but is first “scored” as to relevance then chosen?
The lore entries (and memories) are ‘embedded’ and are represented with an array of floats. The search queries are also embedded, and both search query and the entries are computed with a dot product to get the score. If the search query generated have the other character’s name, it might score the lore entries from that character more likely.
That makes sense; I want to learn more about how the weighting is computed (quite interesting stuff) but dot-producting the arrays to get an overall “this is relevant” score is useful knowledge.
Makes me think that we could weight each individual Lore entry a little more knowing this.
“Tim is the sheriff. Tim used to be the deputy. Tim is a tall guy” might weight-out better than: “Tim is the sheriff, but he used to be the deputy and he’s tall guy”.
That could get the “Tim” weighting a bit higher for that Lore if I’m understanding right.
Thanks!