Architecture

Learn how to build Excel files programmatically with ExcelJS and process them directly in Power Pages through automated workflows.

I check the new path from worker to search so that a long capture is not only processed, but also appears cleanly as one result.

I rebuilt the worker so that one capture produces multiple token chunks, each chunk is embedded individually, and storage in Qdrant remains completely isolated.

The migration to bge-m3 brought an 8192-token limit. Nevertheless, I deliberately set the chunk limit to 1500 tokens to preserve semantic focus.

I had already implemented text splitting, but the context error kept appearing. Only when I accounted for the token budget and special tokens did the cause become clear.

I do not treat the choice of embedding model as an interchangeable environment variable. In Qdrant, it becomes part of every collection. Using different vector spaces locally and in production creates an expensive future migration.