Guides don't trigger Data Bank/RAG

[Thertzlor]

Hello, I have been heavily using this extension to guide my RPs and have recently started to experiment with the Data Bank feature in ST to improve my characters' memories.

However using guided responses/continues/swipes etc. doesn't seem to work with the Data Bank at all, which makes those tools less useful than they should be.
I have a feeling that this is a bug or oversight, because other triggers tend to work:

E.g. If I have a lorebook entry for "greenberries" and I trigger a guided response telling a character to "remember the greenberries", the lorebook entry will be included in the prompt.
But If I have vectorized a document describing "greenberries" vectorized in the Data Bank, the guide will never cause any chunks of it to be included, even if inserting a system message with the exact same wording will insert the relevant parts.

This often leads to strange behavior when a character has no clue about some term in the guide text and just makes something up, only to "remember" what it's supposed to be when their own mention of the term triggers the Data Bank when the next response is sent.

Is it perhaps possible that the guides are inserted after the prompt interceptor for Vector Storage has already run?

[onestardao]

looks like what you’re hitting is not just a “guide vs databank” mismatch, but a sequencing drift. basically the guide gets injected after the vector retrieval layer already ran, so your databank never sees the trigger term in time.

this is the same failure we mapped as ProblemMap No.8 (retriever mismatch). the fix isn’t infra — it’s semantic alignment: making sure guide insertions and retrieval happen in the same semantic pass instead of two disconnected layers.

if you want, I can point you to the checklist we use so you can trace this step-by-step. let me know if you’d like the link.

[Samueras]

Yeah that would be great Thanks. Is am starting to get to the edges of my skills and ever help is very welcome.

[onestardao]

@Samueras @Thertzlor

looks like what you’re hitting maps to Problem Map No.8 – retriever mismatch / sequencing drift. the guides are being injected after the vector store retrieval has already fired, so the databank never sees them in time.

quick tip to test:
before the retriever runs, fold the guide text into the same request object as the user input, so they share the same embedding call. then log ΔS between the guide terms and retrieved spans — if ΔS drops below ~0.45, you’ve got confirmation of the mismatch.

full checklist and step-by-step fix is here:
👉 WFGY Problem Map

[Samueras]

@Thertzlor Can you test something for me? If you make a Custom Guide. Does that get picked up by the Data Bank feature?

[Thertzlor]

@Thertzlor Can you test something for me? If you make a Custom Guide. Does that get picked up by the Data Bank feature?

No, it doesn't seem to matter which kind of guide I use, typed or inserted with the menu. The different modes like response/swipe/continue also don't make a difference.

But I was looking at how the guides are actually inserted, and I get the feeling that this might be a more general issue with how ST handles the chat_metadata.script_injects object.
I'm still testing around but it seems that the built-in Author's Note script that works similarly also doesn't seem to trigger the vector storage.