Historical archives contain rich records of legaland administrative life, but semantic search over these collections remains difficult: keyword search fails when historical vocabulary differs from modern queries, and no labeled training data exists for supervised approaches. We address this challenge in the context of 19th-century Swedish archival documents from the Swedish National Archives, where HTR transcription noise and centuries of linguistic drift compound the difficulty further.

The key contribution is an N-to-N synthetic querygeneration strategy, designed around the insightthat queries should reflect how historians actually search: blind searches that span multiple related records, not paraphrases of individual documents. Combined with LoRA fine-tuning and cumulative domain expansion across four document types, the approach achieves consistent improvements acrossall evaluation sets. The results suggest that effective domain adaptation is achievable for this challenging historical retrieval setting, and that data strategy is the critical determinant of success.