Local-First AI Research: What Stays on Your Computer?
A local project can still send text to cloud models. Check generation, embeddings, external tools and licensing separately before using sensitive research material.
Local-first describes where your working files live. It does not, by itself, tell you where AI processing happens. If a research app keeps a PDF on your disk but sends its paragraphs to a cloud model, the file stays local while the text crosses the network.
For Scholardo, check four choices separately: the conversation model, background and editing models, document embeddings, and any external tools or CLIs you connect. You can use local generation and local embeddings, or choose cloud services with data handling appropriate to your project.
Start with the task, not the label
Suppose you want to compare three unpublished manuscripts. Before asking for a summary, trace the task:
- Where are the manuscripts stored?
- Which service will read their text to produce the answer?
- Has their text already been sent somewhere to build a search index?
- Can the agent run commands or call external services beyond that model request?
A setting for one step does not settle the others. A local conversation model does not turn a remote embedding endpoint into a local one. A project-folder check on a file tool does not limit every command a terminal agent can run.
Generation and embeddings are different
Generation produces the response: a summary, a draft, a proposed concept or an edit. When a configured cloud model does that work, it receives the task text. That can include passages from your research files. Background tasks you enable can use a model too; closing the chat composer is not a general switch for all model work.
Embeddings support semantic search. An embedding service processes chunks of a document to build vectors for retrieval. The vectors may be stored locally even when the text was processed remotely. Check the embedding provider as well as the conversation model.
In Scholardo, model providers and the default model live under Settings → AI Agent → Models. Embeddings have their own settings under Settings → Embedding. Local Ollama services can handle either task, but each needs its own appropriate model and configuration. Without an embedding provider, keyword search remains available.
The data-flow guide maps these requests, and the search guide explains indexing and consent.
File tools and command permissions are different
Scholardo's file tools check the current project and linked external sources. That is useful for keeping a tool request within its intended material. It is not an operating-system sandbox around the whole agent.
A command, or a third-party CLI running in a terminal, has its own permissions under your user account. Review those settings if you use a CLI. Putting a confidential file outside the project is not enough to prove that every program running under your account cannot read it.
This distinction matters even with a local model. Local inference answers where the model runs; it does not answer every question about command execution or external tool access.
Offline work has conditions too
Reading saved files, citation rendering and local processing do not require a cloud model. Fetching a paper, refreshing a feed or opening a web page requires a connection to that source.
Licensing is another separate request. Scholardo requires online activation and periodically refreshes its license. Fully offline, it keeps working for at least two weeks before becoming read-only until it connects again. That condition still applies when generation and embeddings are local. See license and offline behavior.
A practical configuration check
Before adding sensitive material, use a non-sensitive sample and inspect the selected generation and embedding services. Review background work and external tools. Confirm that any linked folders contain material you intend to make available. If you use a CLI, review its permissions separately.
This is a configuration check, not a certification that a study complies with an institution's rules. Your institution and the relevant service terms determine whether a particular use is appropriate. Local storage alone cannot answer that decision.
For a first task with ordinary, non-sensitive material, follow Your First Research Session. For the full account of Scholardo's data handling, read the Privacy Policy.