Semantic search

Semantic search finds text by meaning, not by exact words. Alisio ships the shared capability; the engine that does the work comes from a plugin you choose to turn on.

What it is

A normal search looks for the words you typed. A semantic search looks for the idea behind them: a search for "budget meeting" can find a note titled "quarterly financial review" even though they share no words. Depending on the model the engine uses, it can also work across languages, so a question in Spanish can find a document written in English.

What it enables

Semantic search is a building block. Plugins that use it can offer:

• Chat with your own documents, with answers that cite the passages they came from (a research notebook over your sources).
• Smarter memory recall, finding what you saved by meaning and not only by keywords.
• Related files and notes, so you can find what is close to what you are reading.
• Deduplicating and grouping similar content, such as near-identical documents or repeated notes.

How it works in Alisio

• Alisio defines the capability. It provides the shared contract and the service that plugins use, and nothing heavier, so the app stays small.
• A provider plugin supplies the engine: an embedding model (turns text into numbers that capture meaning) and a vector store (keeps and searches them).
• Installing a plugin never turns it on. You activate it, and your choice is saved in your global configuration. Project configuration cannot activate a provider.
• /semantic shows the status: which providers are registered and active, the model and the health of each one.
• It can run locally. Whether anything leaves your computer depends on the provider you choose.

Current status

Install a provider to turn it on

Alisio 0.11.7 ships the capability, not the engine, so the app stays small. The provider is @alisio/plugin-semantic from the alisio-plugins catalog: it runs a local LanceDB store with the multilingual-e5-small model and includes a Notebook app to chat with your sources, with cited answers. Install it like any other plugin and activate it when it asks. Without an active provider, plugins fall back to their own behavior, for example keyword search.

Privacy and limits

• Vectors and indexes stay wherever the provider keeps them. Check what your provider does before you index private documents.
• Changing the model requires indexing again: vectors from different models are not comparable, so existing collections are rejected until they are rebuilt.
• Each plugin has its own namespace of collections. Namespaces prevent accidental mixing; they are not a security sandbox, because plugins are trusted code.
• You can turn the feature off with semantic.enabled.

For plugin developers

Plugins can consume the capability through context.semantic or provide an engine through api.semantic. Both members are optional, so feature-detect them and keep a fallback. See Semantic search providers and consumers for the API and the semantic configuration key for the settings.