First, the process
A local model can support document search, classification or drafting responses. The choice starts from the task, data sensitivity, latency and the team’s ability to operate the system.
Access follows the user’s rights
Connecting documents to AI requires preserving permissions. Indexing, information retrieval, logs and backups are part of the same architecture. Local hosting does not automatically eliminate risks.
Quality is checked on relevant examples
We build an evaluation set and compare responses, correct refusals and cited sources. We choose between configuration, retrieval from documents and model fine-tuning based on results.
Operation matters after the pilot
Hardware sizing, updates, monitoring and rollback are part of the project. In large organisations, including the banking sector, responsibilities and acceptance criteria must be documented before scaling.