Vector search with Basis

Prerequisites

Use an authorized local Weave source checkout. Add weave-sdk as a path dependency to libs/weave-sdk; its manifest version is 1.1.0. This avoids assuming a public registry artifact. The snippets use the current SDK methods in libs/weave-sdk/src/node.rs.

[dependencies]
weave-sdk = { path = "../weave/libs/weave-sdk" }
tokio = { version = "1", features = ["rt-multi-thread", "macros"] }

Adjust the path from your application's Cargo.toml. Use an application-owned local data directory. The sample opens local stores and does not start networking.

Add uuid = { version = "1", features = ["v4"] } to dependencies.

Verify vector mechanics

The fixed vectors below are numeric fixtures. They demonstrate nearest-neighbor storage, not semantic quality. Production ingestion and query vectors must use the same model, dimensionality and normalization.

use weave_sdk::prelude::*;

#[tokio::main]
async fn main() -> WeaveResult<()> {
    let node = WeaveNode::builder()
        .namespace("docs-example")
        .identifier("vectors")
        .storage_dir("./data/vectors")
        .build()
        .await?;
    node.open_basis("documents").await?;
    let document_id = uuid::Uuid::new_v4();
    let embedding = vec![1.0_f32, 0.0, 0.0];
    node.basis_add("documents", document_id, &embedding).await?;
    let hits = node.basis_search("documents", &embedding, 1).await?;
    for (id, distance) in hits {
        println!("{id}: {distance}");
    }
    Ok(())
}

Connect documents to embeddings

Store source text in Strand or Locus and retain its stable ID alongside each vector. Replace the fixture with a configured embedding provider and record model/version/dimensions with the collection. Resolve returned UUIDs through your application index. Rebuild vectors after incompatible model changes; do not silently mix embeddings.

Basis reference covers exact source declarations; Gnosis can hold related graph metadata. Neither component supplies the embedding model.