Your RAG over PDF reports handles prose well but gets n...

Do not start by changing the embedding model or adding a reranker, as they cannot fix issues where the necessary context (column names) has been stripped away. Inspect the retrieved chunks to see if the table structure is being preserved. If the header and data rows are separated, the retriever will fail to associate the query vocabulary with the relevant data rows. The solution involves ensuring tables are ingested in a way that preserves the relationship between headers and values.