Semantic search

Two of the three bands in search are words, and need nothing. The third finds notes that are about what you typed, and for that your notes have to be read by a model.

Out of the box that model is small, runs on your machine, and is fetched the first time it is wanted. No key, no account, and nothing you write leaves the computer.

Everything here is indexing.embedding in the settings file.

Turning it off

Search is then names and words alone, nothing is fetched and nothing is asked of a network:

{ "indexing": { "embedding": { "indexing": { "use": "" } } } }

Handing it to a service

Anything speaking the usual embeddings request will do:

{
  "indexing": {
    "embedding": {
      "indexing": {
        "use": "service",
        "service": {
          "base_url": "https://api.openai.com/v1",
          "name": "text-embedding-3-small",
          "key_env": "OPENAI_API_KEY"
        }
      }
    }
  }
}

The key is read from the environment variable you name. numen never writes a key back into the file.

Two places, because the costs are opposite

indexing is where vectors are made while your vault is read. query is where the vector for a question is made. They are separate settings because they are not the same job:

• Filling the index is one pass over everything you have. A service does in an hour what a laptop does in a day.
• Asking a question is twenty words, all day long. Your machine answers in milliseconds where a network is a round trip — and answers on a train, with no network at all.

So a good arrangement is often: a service fills the index, your machine answers the questions. The two must be the same model, and numen checks that they are by embedding the same short text in both at startup:

{
  "indexing": {
    "embedding": {
      "model": { "name": "bge-m3", "dimensions": 1024, "max_tokens": 512, "pooling": "head" },
      "indexing": {
        "use": "service",
        "service": {
          "base_url": "https://openrouter.ai/api/v1",
          "name": "baai/bge-m3",
          "key_env": "OPENROUTER_API_KEY"
        }
      },
      "query": { "use": "local", "local": { "name": "BAAI/bge-m3", "download": true } }
    }
  }
}

The three names differ on purpose: model.name is what you call the model, and the other two are what each place calls it.

query left unset asks the way the vault was indexed.

Changing the model

name, dimensions, max_tokens and pooling are what a vector is kept under, so changing any of them means your vault is read again. Nothing is thrown away: setting the old ones back finds the old work exactly where it was.

pooling is the one that goes wrong quietly. A model gathers what a text says either into the token that opens it or across all of them, and taken the wrong way it answers with vectors that mean nothing — near nothing, with no error anywhere. mean is the E5 family, sentence-transformers, and most of what is published; head is BGE, bge-m3 included.

The rest

floor is how near a question a passage has to stand to count as an answer. Zero takes what the search was built against. Where a model puts two pieces of text is a fact about that model, so a model changed is a floor worth measuring again.

local.dir names a folder holding a model already on the machine, which is what an installation with no network uses. local.file picks a particular build inside a repository — how a quantised model is run in place of the full one.

The complete list is on every setting.