"Phantom prompt injection"

Al borrar documentos de la KB, los embeddings persisten en el vector store y siguen siendo servidos a los agentes. Esto tiene implicaciones de privacidad (contenido de un agente eliminado puede filtrarse a otros) y de comportamiento (los agentes reciben contexto “fantasma” que no existe en la UI).

Documentos borrados: plantilla-entrevista.md, Carmen Moreira.md, A2 - Callum Redactor M1.md, A2A - Callum Alpha.md
Síntoma: agentes siguen viendo contenido de esos documentos vía RAG después de borrarlos
Workspace ID: 3afbaed5-983c-4ba9-978b-99398cd3901a

Hello @rcasanuevau,

We could not find evidence that deleted knowledge base document embeddings were still being retrieved in your workspace. Could you please provide the Session IDs where you noticed this behavior occurring?

We found one run which referenced your Callum documents after deletion but that was due to the document being retrieved pre-deletion which kept the contents in the conversation context, not because the embeddings persisted after deletion.

Please let us know any relevant information so that we can resolve this issue for you!