LOCAL AI / CONTINUITY
Keep the model.
Recover the work as well.
A model file can be very large and give the impression that the essentials are backed up. Yet the documents, decisions and settings that make the assistant useful often live elsewhere. Continuity is planned at the level of the whole service.
Written by Saphan · 7 September 2026
What the weights contain, and what they do not replace
The weights are parameters produced by training. They are not a faithful, restorable register of your invoices, procedures or cases. Even a model adapted to your business does not replace keeping the originals and the validated decisions.
You also need to identify the model used precisely: version, format, any adaptation and terms of use. An identical commercial name is not enough to recover the same installation.
Assemble a coherent restore set
- Execution: model, engine, dependencies and the settings needed.
- Application: code, instructions, connectors, configuration and acceptance tests.
- Business: documents, case database, decisions, permissions and versions.
- Access: a separate procedure to restore the authorized identities and secrets, without spreading them through ordinary copies.
A search index can be rebuilt if the sources and the processing chain are kept and compatible. Rebuilding nevertheless takes time. You need to decide which items to back up directly and which to regenerate, then measure the real time it takes.
Date the copy and define the work to recover
A copy of documents and a database saved at different times can describe contradictory states. Define a consistent recovery point, the acceptable loss of work and the target downtime. These objectives must be tested in an exercise, not just written down in a file.
Fictional example: after a restore, the assistant finds an order but ignores its recent validation. The model works; the business service remains incomplete. The test must therefore check the state of a case, not just get a reply in the chat.
Restore separately, then have someone take over
Saphan can prepare an isolated trial with authorized data: startup, restore, controlled reconnection and an acceptance run. A second person follows the procedure, checks the rights and notes what is missing. The result sought is a usable, dated version, with its recovery limits stated explicitly.