diff --git a/.github/workflows/rust-vector-tests.yml b/.github/workflows/rust-vector-tests.yml
new file mode 100644
index 0000000000..ca497ee71e
--- /dev/null
+++ b/.github/workflows/rust-vector-tests.yml
@@ -0,0 +1,76 @@
+name: "Rust Vector Tests"
+
+on:
+  pull_request:
+    branches: [ master ]
+    paths:
+      - 'seaweed-vector/**'
+      - 'seaweed-worker/**'
+      - 'seaweed-common/**'
+      - 'weed/pb/plugin.proto'
+      - '.github/workflows/rust-vector-tests.yml'
+  push:
+    branches: [ master, main ]
+    paths:
+      - 'seaweed-vector/**'
+      - 'seaweed-worker/**'
+      - 'seaweed-common/**'
+      - 'weed/pb/plugin.proto'
+      - '.github/workflows/rust-vector-tests.yml'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || (github.event_name == 'push' && github.ref) || github.run_id }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' || (github.event_name == 'push' && !startsWith(github.ref, 'refs/tags/')) }}
+
+permissions:
+  contents: read
+
+jobs:
+  rust-vector-build:
+    name: Rust Vector Build and Unit Tests
+    runs-on: ubuntu-24.04
+    timeout-minutes: 45
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v7
+        with:
+          persist-credentials: false
+
+      - name: Install Rust toolchain
+        uses: dtolnay/rust-toolchain@stable
+
+      - name: Fingerprint build toolchain
+        id: toolchain
+        run: echo "fingerprint=$(getconf GNU_LIBC_VERSION | tr ' ' '-')-rustc-$(rustc -V | awk '{print $2}')" >> "$GITHUB_OUTPUT"
+
+      - name: Cache cargo registry and target
+        if: runner.environment == 'github-hosted'
+        uses: actions/cache@v6
+        with:
+          path: |
+            ~/.cargo/registry
+            ~/.cargo/git
+            seaweed-vector/target/release
+          key: rust-vector-${{ steps.toolchain.outputs.fingerprint }}-${{ hashFiles('seaweed-vector/Cargo.lock') }}
+          restore-keys: |
+            rust-vector-${{ steps.toolchain.outputs.fingerprint }}-
+
+      # lance's build scripts compile their own protos and look for a protoc.
+      # Point them at the one protoc-bin-vendored ships, so no job depends on a
+      # system package and every build sees the same version.
+      - name: Use the vendored protoc
+        run: |
+          cd seaweed-vector
+          cargo fetch
+          version=$(awk '/^name = "protoc-bin-vendored-linux-x86_64"$/{found=1; next} found && /^version = /{gsub(/"/,"",$3); print $3; exit}' Cargo.lock)
+          test -n "$version" || { echo "protoc-bin-vendored-linux-x86_64 is not in Cargo.lock" >&2; exit 1; }
+          protoc=$(find ~/.cargo/registry/src -path "*protoc-bin-vendored-linux-x86_64-$version/bin/protoc" | head -1)
+          test -x "$protoc" || { echo "no vendored protoc $version in the registry" >&2; exit 1; }
+          echo "PROTOC=$protoc" >> "$GITHUB_ENV"
+
+      - name: Build the vector crates
+        run: cd seaweed-vector && cargo build --release
+
+      - name: Run unit tests
+        run: cd seaweed-vector && cargo test --release --workspace
diff --git a/seaweed-vector/Cargo.lock b/seaweed-vector/Cargo.lock
index c4780cdd26..879a1c4fbb 100644
--- a/seaweed-vector/Cargo.lock
+++ b/seaweed-vector/Cargo.lock
@@ -6457,7 +6457,9 @@ version = "0.1.0"
 dependencies = [
  "anyhow",
  "arrow-array",
+ "arrow-cast",
  "arrow-schema",
+ "clap",
  "futures",
  "lance",
  "lance-index",
@@ -6469,6 +6471,7 @@ dependencies = [
  "tempfile",
  "tokio",
  "tracing",
+ "tracing-subscriber",
  "weed-lance-worker",
 ]
 
diff --git a/seaweed-vector/crates/core/Cargo.toml b/seaweed-vector/crates/core/Cargo.toml
index 4241f4a136..7888bbe51b 100644
--- a/seaweed-vector/crates/core/Cargo.toml
+++ b/seaweed-vector/crates/core/Cargo.toml
@@ -8,6 +8,10 @@ description = "Vector query engine over SeaweedFS Lance table buckets"
 [lib]
 name = "weed_vector_core"
 
+[[bin]]
+name = "weed-vector"
+path = "src/main.rs"
+
 [dependencies]
 # The catalog client and dataset-opening helper the maintenance worker uses.
 # A path dependency because the two trees are separate cargo workspaces with
@@ -18,14 +22,17 @@ lance-index = "10"
 lance-linalg = "10"
 lance-table = "10"
 arrow-array = "58"
+arrow-cast = "58"
 arrow-schema = "58"
 roaring = "0.11"
 anyhow.workspace = true
+clap = { version = "4", features = ["derive", "env"] }
 futures = "0.3"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 tokio.workspace = true
 tracing.workspace = true
+tracing-subscriber.workspace = true
 
 [dev-dependencies]
 tempfile = "3"
diff --git a/seaweed-vector/crates/core/src/coverage.rs b/seaweed-vector/crates/core/src/coverage.rs
index 02e537922f..4dd68620a6 100644
--- a/seaweed-vector/crates/core/src/coverage.rs
+++ b/seaweed-vector/crates/core/src/coverage.rs
@@ -8,7 +8,7 @@ use lance::index::DatasetIndexExt;
 use roaring::RoaringBitmap;
 
 /// How much of the dataset one named index covers at this version.
-#[derive(Debug)]
+#[derive(Debug, serde::Serialize)]
 pub struct IndexCoverage {
     pub index_name: String,
     /// Physical index segments: the IndexMetadata deltas sharing the name.
diff --git a/seaweed-vector/crates/core/src/lib.rs b/seaweed-vector/crates/core/src/lib.rs
index c4842ddbae..e4cb1453d2 100644
--- a/seaweed-vector/crates/core/src/lib.rs
+++ b/seaweed-vector/crates/core/src/lib.rs
@@ -10,6 +10,8 @@ pub mod query;
 pub mod table;
 #[cfg(test)]
 mod testing;
+#[cfg(test)]
+mod verify;
 
 pub use coverage::IndexCoverage;
 pub use query::{Hit, Query};
diff --git a/seaweed-vector/crates/core/src/main.rs b/seaweed-vector/crates/core/src/main.rs
new file mode 100644
index 0000000000..879dfb362a
--- /dev/null
+++ b/seaweed-vector/crates/core/src/main.rs
@@ -0,0 +1,209 @@
+use anyhow::{Context, Result, bail};
+use clap::{Parser, Subcommand};
+
+use weed_lance_worker::catalog::{NamespaceClient, parse_id};
+use weed_lance_worker::dataset::{self, FallbackOptions};
+use weed_vector_core::coverage;
+use weed_vector_core::query::{self, Query};
+use weed_vector_core::table::VectorTable;
+
+/// Mirrors `weed-worker`'s flags where the two touch the same thing: the
+/// namespace is WEED_LANCE_NAMESPACE and the fallback storage credentials are
+/// the same env vars.
+#[derive(Parser, Debug)]
+#[command(name = "weed-vector", about = "SeaweedFS vector query spike")]
+struct Args {
+    /// Lance namespace the tables live in.
+    #[arg(
+        long,
+        default_value = "http://localhost:9101",
+        env = "WEED_LANCE_NAMESPACE"
+    )]
+    namespace: String,
+
+    /// Storage credentials to use where the namespace vends none.
+    #[arg(long, env = "WEED_S3_ACCESS_KEY")]
+    access_key: Option<String>,
+
+    #[arg(long, env = "WEED_S3_SECRET_KEY")]
+    secret_key: Option<String>,
+
+    #[command(subcommand)]
+    command: Command,
+}
+
+#[derive(Subcommand, Debug)]
+enum Command {
+    /// A table's location, version, vector layout and index coverage.
+    Describe {
+        /// Delimiter-joined table id, e.g. bucket$table.
+        id: String,
+    },
+    /// Which fragments each vector index covers.
+    Coverage { id: String },
+    /// A top-K query over the whole dataset.
+    Query {
+        id: String,
+        /// Comma-separated floats, e.g. --vector 0.1,0.2,0.3,0.4
+        #[arg(long)]
+        vector: String,
+        #[arg(long, default_value = "10")]
+        k: usize,
+        /// Lance SQL filter over the metadata columns.
+        #[arg(long)]
+        filter: Option<String>,
+        /// l2, cosine or dot.
+        #[arg(long, default_value = "l2")]
+        metric: String,
+        /// Exact scan, no index.
+        #[arg(long)]
+        flat: bool,
+        /// Filter the top-K rows after ranking (Lance's default); off keeps
+        /// the filter before the ANN search.
+        #[arg(long)]
+        postfilter: bool,
+        /// Pin the query to this committed version.
+        #[arg(long)]
+        version: Option<u64>,
+        #[arg(long)]
+        vector_column: Option<String>,
+        #[arg(long)]
+        key_column: Option<String>,
+        #[arg(long)]
+        nprobes: Option<usize>,
+    },
+    /// Fetch rows by key.
+    Get { id: String, keys: Vec<String> },
+}
+
+/// `id` is a delimiter-joined table id; when it looks like a URI (`s3://...`,
+/// `/path/...`, `file://...`) the dataset is opened directly, skipping the
+/// namespace — useful for a spike run against a local dataset.
+async fn open_dataset(
+    client: &NamespaceClient,
+    id: &str,
+    fallback: &FallbackOptions,
+) -> Result<lance::Dataset> {
+    if id.contains("://") || id.starts_with('/') {
+        let mut builder = lance::dataset::builder::DatasetBuilder::from_uri(id);
+        if !fallback.is_empty() {
+            builder = builder.with_storage_options(fallback.clone());
+        }
+        return builder.load().await.with_context(|| format!("open {id}"));
+    }
+    let table = dataset::open(client, &parse_id(id), fallback).await?;
+    Ok(table.dataset)
+}
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    weed_lance_worker::tls::install_default_crypto_provider();
+    tracing_subscriber::fmt()
+        .with_env_filter(
+            tracing_subscriber::EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new("info")),
+        )
+        .init();
+    let args = Args::parse();
+
+    let mut fallback = FallbackOptions::new();
+    if let (Some(access), Some(secret)) = (args.access_key, args.secret_key) {
+        fallback.insert("aws_access_key_id".to_string(), access);
+        fallback.insert("aws_secret_access_key".to_string(), secret);
+    }
+    let client = NamespaceClient::new(args.namespace).with_fallback(&fallback);
+
+    match args.command {
+        Command::Describe { id } => {
+            let dataset = open_dataset(&client, &id, &fallback).await?;
+            let schema: arrow_schema::Schema = dataset.schema().into();
+            let vector = VectorTable::resolve(&dataset, None, None).ok();
+            let coverages = match &vector {
+                Some(v) => coverage::coverage(&dataset, &v.vector_column).await?,
+                None => Vec::new(),
+            };
+            println!(
+                "{}",
+                serde_json::json!({
+                    "id": id,
+                    "uri": dataset.uri(),
+                    "version": dataset.version().version,
+                    "fragments": dataset.get_fragments().len(),
+                    "rows": dataset.count_rows(None).await?,
+                    "schema": schema.fields().iter().map(|f| serde_json::json!({
+                        "name": f.name(),
+                        "type": f.data_type().to_string(),
+                        "nullable": f.is_nullable(),
+                    })).collect::<Vec<_>>(),
+                    "vector": vector.map(|v| serde_json::json!({
+                        "vector_column": v.vector_column,
+                        "dimension": v.dimension,
+                        "key_column": v.key_column,
+                        "metadata_columns": v.metadata_columns,
+                    })),
+                    "coverage": coverages,
+                })
+            );
+        }
+        Command::Coverage { id } => {
+            let dataset = open_dataset(&client, &id, &fallback).await?;
+            let vector = VectorTable::resolve(&dataset, None, None)?;
+            let coverages = coverage::coverage(&dataset, &vector.vector_column).await?;
+            println!("{}", serde_json::to_string_pretty(&coverages)?);
+        }
+        Command::Get { id, keys } => {
+            let dataset = open_dataset(&client, &id, &fallback).await?;
+            let vector = VectorTable::resolve(&dataset, None, None)?;
+            for batch in query::get(&dataset, &vector, &keys).await? {
+                println!("{}", arrow_cast::pretty::pretty_format_batches(&[batch])?);
+            }
+        }
+        Command::Query {
+            id,
+            vector,
+            k,
+            filter,
+            metric,
+            flat,
+            postfilter,
+            version,
+            vector_column,
+            key_column,
+            nprobes,
+        } => {
+            let opened = open_dataset(&client, &id, &fallback).await?;
+            let dataset = match version {
+                Some(v) => opened.checkout_version(v).await?,
+                None => opened,
+            };
+            let vector_table =
+                VectorTable::resolve(&dataset, vector_column.as_deref(), key_column.as_deref())?;
+            let vector: Vec<f32> = vector
+                .split(',')
+                .map(|part| part.trim().parse::<f32>())
+                .collect::<std::result::Result<_, _>>()
+                .context("--vector takes comma-separated floats")?;
+            let distance = match metric.as_str() {
+                "l2" | "euclidean" => lance_linalg::distance::DistanceType::L2,
+                "cosine" => lance_linalg::distance::DistanceType::Cosine,
+                "dot" | "dotproduct" => lance_linalg::distance::DistanceType::Dot,
+                other => bail!("unknown metric {other}: l2, cosine or dot"),
+            };
+            let query = Query {
+                vector,
+                k,
+                filter,
+                postfilter,
+                distance,
+                nprobes,
+            };
+            let hits = if flat {
+                query::flat(&dataset, &vector_table, &query).await?
+            } else {
+                query::ann(&dataset, &vector_table, &query).await?
+            };
+            println!("{}", serde_json::to_string_pretty(&hits)?);
+        }
+    }
+    Ok(())
+}
diff --git a/seaweed-vector/crates/core/src/query.rs b/seaweed-vector/crates/core/src/query.rs
index fb917adfec..420afb69fb 100644
--- a/seaweed-vector/crates/core/src/query.rs
+++ b/seaweed-vector/crates/core/src/query.rs
@@ -30,7 +30,7 @@ pub struct Query {
 }
 
 /// One row of a top-K answer.
-#[derive(Debug, Clone)]
+#[derive(Debug, Clone, serde::Serialize)]
 pub struct Hit {
     /// The key column value, or the row id when the table has no key column.
     pub key: String,
diff --git a/seaweed-vector/crates/core/src/verify.rs b/seaweed-vector/crates/core/src/verify.rs
new file mode 100644
index 0000000000..2ba0b91b8d
--- /dev/null
+++ b/seaweed-vector/crates/core/src/verify.rs
@@ -0,0 +1,201 @@
+//! The spike's exit gate: the engine's answers must equal what a bare Lance
+//! `scan().nearest()` produces on the same pinned version, and no eligible
+//! row — indexed or not — may be missing.
+
+use arrow_array::{Float32Array, UInt64Array};
+use futures::TryStreamExt;
+use lance::Dataset;
+use lance::index::DatasetIndexExt;
+use lance::index::vector::VectorIndexParams;
+use lance_index::IndexType;
+use lance_index::optimize::OptimizeOptions;
+use lance_linalg::distance::DistanceType;
+use tempfile::TempDir;
+
+use crate::coverage::coverage;
+use crate::query::{self, Query};
+use crate::table::VectorTable;
+use crate::testing::{self, DIM};
+
+async fn indexed_dataset(dir: &TempDir) -> Dataset {
+    let mut dataset = testing::write_dataset(dir, &testing::vectors(64, DIM), &["a"; 64]).await;
+    let params = VectorIndexParams::ivf_flat(4, DistanceType::L2);
+    dataset
+        .create_index(
+            &["vector"],
+            IndexType::Vector,
+            Some("vidx".to_string()),
+            &params,
+            true,
+        )
+        .await
+        .unwrap();
+    dataset.checkout_latest().await.unwrap();
+    dataset
+}
+
+fn query_for(vector: Vec<f32>, k: usize) -> Query {
+    Query {
+        vector,
+        k,
+        filter: None,
+        postfilter: false,
+        distance: DistanceType::L2,
+        nprobes: Some(4),
+    }
+}
+
+/// The reference path: a bare scanner over the same dataset handle.
+async fn reference(dataset: &Dataset, vector: &[f32], k: usize) -> Vec<(u64, f32)> {
+    let mut scanner = dataset.scan();
+    scanner
+        .nearest("vector", &Float32Array::from(vector.to_vec()), k)
+        .unwrap()
+        .distance_metric(DistanceType::L2)
+        .nprobes(4)
+        .with_row_id();
+    let mut stream = scanner.try_into_stream().await.unwrap();
+    let mut rows = Vec::new();
+    while let Some(batch) = stream.try_next().await.unwrap() {
+        let ids = batch
+            .column_by_name("_rowid")
+            .unwrap()
+            .as_any()
+            .downcast_ref::<UInt64Array>()
+            .unwrap();
+        let distances = batch
+            .column_by_name("_distance")
+            .unwrap()
+            .as_any()
+            .downcast_ref::<Float32Array>()
+            .unwrap();
+        for i in 0..batch.num_rows() {
+            rows.push((ids.value(i), distances.value(i)));
+        }
+    }
+    rows
+}
+
+async fn assert_matches_reference(dataset: &Dataset, vector: &[f32], k: usize) {
+    let table = VectorTable::resolve(dataset, None, None).unwrap();
+    let hits = query::ann(dataset, &table, &query_for(vector.to_vec(), k))
+        .await
+        .unwrap();
+    let expected = reference(dataset, vector, k).await;
+    assert_eq!(hits.len(), expected.len());
+    for (hit, (row_id, distance)) in hits.iter().zip(expected.iter()) {
+        assert_eq!(hit.row_id, *row_id);
+        assert!((hit.distance - distance).abs() < 1e-6, "{}", hit.key);
+    }
+}
+
+#[tokio::test]
+async fn ann_matches_lance_reference() {
+    let dir = TempDir::new().unwrap();
+    let dataset = indexed_dataset(&dir).await;
+    assert_matches_reference(&dataset, &testing::vectors(1, DIM)[0], 10).await;
+}
+
+#[tokio::test]
+async fn ann_matches_reference_over_unindexed_fragments() {
+    let dir = TempDir::new().unwrap();
+    let mut dataset = indexed_dataset(&dir).await;
+    testing::append_dataset(&mut dataset, &testing::vectors(8, DIM), &["a"; 8]).await;
+
+    let cov = &coverage(&dataset, "vector").await.unwrap()[0];
+    assert_eq!(cov.unindexed_fragments.len(), 1);
+    assert_eq!(cov.unindexed_rows, 8);
+
+    for vector in testing::vectors(4, DIM) {
+        assert_matches_reference(&dataset, &vector, 10).await;
+    }
+}
+
+#[tokio::test]
+async fn ann_matches_reference_across_segments() {
+    let dir = TempDir::new().unwrap();
+    let mut dataset = indexed_dataset(&dir).await;
+    testing::append_dataset(&mut dataset, &testing::vectors(32, DIM), &["a"; 32]).await;
+    dataset
+        .optimize_indices(&OptimizeOptions::append())
+        .await
+        .unwrap();
+
+    let cov = &coverage(&dataset, "vector").await.unwrap()[0];
+    assert_eq!(cov.segments, 2);
+    assert!(cov.unindexed_fragments.is_empty());
+
+    for vector in testing::vectors(4, DIM) {
+        assert_matches_reference(&dataset, &vector, 10).await;
+    }
+}
+
+#[tokio::test]
+async fn deletes_are_not_returned() {
+    let dir = TempDir::new().unwrap();
+    let mut dataset = indexed_dataset(&dir).await;
+    dataset.delete("key = 'k0'").await.unwrap();
+    assert_matches_reference(&dataset, &testing::vectors(1, DIM)[0], 10).await;
+}
+
+/// Writes a small indexed dataset with unindexed rows to
+/// $WEED_VECTOR_SPIKE_DIR for exercising the `weed-vector` CLI:
+///
+///   WEED_VECTOR_SPIKE_DIR=/tmp/vds cargo test make_spike_dataset -- --ignored
+///   weed-vector describe /tmp/vds
+///   weed-vector coverage /tmp/vds
+///   weed-vector query /tmp/vds --vector 0,0,0,0 --k 5
+#[tokio::test]
+#[ignore]
+async fn make_spike_dataset() {
+    let dir = std::env::var("WEED_VECTOR_SPIKE_DIR").expect("set WEED_VECTOR_SPIKE_DIR");
+    let mut dataset = testing::write_dataset_at(&dir, &testing::vectors(64, DIM), &["a"; 64]).await;
+    let params = VectorIndexParams::ivf_flat(4, DistanceType::L2);
+    dataset
+        .create_index(
+            &["vector"],
+            IndexType::Vector,
+            Some("vidx".to_string()),
+            &params,
+            true,
+        )
+        .await
+        .unwrap();
+    dataset.checkout_latest().await.unwrap();
+    testing::append_dataset(&mut dataset, &testing::vectors(8, DIM), &["b"; 8]).await;
+}
+
+#[tokio::test]
+async fn query_at_pinned_version_survives_maintenance() {
+    let dir = TempDir::new().unwrap();
+    let mut dataset = indexed_dataset(&dir).await;
+    let pinned = dataset
+        .checkout_version(dataset.version().version)
+        .await
+        .unwrap();
+
+    testing::append_dataset(&mut dataset, &testing::vectors(8, DIM), &["b"; 8]).await;
+    dataset
+        .optimize_indices(&OptimizeOptions::default())
+        .await
+        .unwrap();
+
+    assert_eq!(pinned.count_rows(None).await.unwrap(), 64);
+    let vector = testing::vectors(1, DIM)[0].clone();
+    let hits = query::ann(
+        &pinned,
+        &VectorTable::resolve(&pinned, None, None).unwrap(),
+        &query_for(vector.clone(), 10),
+    )
+    .await
+    .unwrap();
+    let expected = reference(&pinned, &vector, 10).await;
+    assert_eq!(hits.len(), expected.len());
+    assert!(hits.iter().all(|h| h.row_id < 64));
+    for (hit, (row_id, distance)) in hits.iter().zip(expected.iter()) {
+        assert_eq!(hit.row_id, *row_id);
+        assert!((hit.distance - distance).abs() < 1e-6);
+    }
+
+    assert_matches_reference(&dataset, &vector, 10).await;
+}