Configure a custom retriever in AI Models, formerly Einstein Studio, with a dynamic filter on the policy version metadata field, populated at runtime from the customer’s Contract record in CRM, ensuring retrieval is deterministically scoped to the document version applicable to that customer’s agreement before similarity ranking is applied.

Add the policy version identifier as a prepend field in the search index so every chunk is tagged with its applicable version at indexing time, enabling the large language model (LLM) to identify and select the correct version during response generation.

Create a separate data stream for each policy version and configure the agent to query the appropriate unstructured data model object (UDMO) corresponding to the customer’s contract date, isolating each version’s vectors from the others.