feat: add support for ngram indices (#3468)

Ngram indices are indices that can speed up various string filters. To
start with they will be able to speed up `contains(col, 'substr')`
filters. They work by creating a bitmap for each ngram (short sequence
of characters) in a value. For example, consider an index of 1-grams.
This would create a bitmap for each letter of the alphabet. Then, at
query time, we can use this to narrow down which strings could
potentially satisfy the query.

This is the first scalar index that requires a "recheck" step. It
doesn't tell us exactly which rows satisfy the query. It only narrows
down the list. Other indices that might behave like this are bloom
filters and zone maps. This means that we need to still apply the filter
on the results of the index search. A good portion of this PR is adding
support for this concept into the scanner.

---------

Co-authored-by: Will Jones <willjones127@gmail.com>

Author: westonpace

.typos.toml

@@ -1,3 +1,6 @@
+[default]
+extend-ignore-re = ["(?Rm)^.*(#|//)\\s*spellchecker:disable-line$"]
+
 [default.extend-words]
 DNE = "DNE"
 arange = "arange"
@@ -7,4 +10,5 @@ abd = "abd"
 afe = "afe"
 
 [files]
-extend-exclude = ["notebooks/*.ipynb"]
+extend-exclude = ["notebooks/*.ipynb"]
+# If a line ends with # or // and has spellchecker:disable-line, ignore it

protos/table.proto

@@ -361,4 +361,5 @@ message BTreeIndexDetails {}
 message BitmapIndexDetails {}
 message LabelListIndexDetails {}
 message InvertedIndexDetails {}
+message NGramIndexDetails {}
 message VectorIndexDetails {}

python/python/lance/dataset.py

@@ -1494,6 +1494,7 @@ def create_scalar_index(
             Literal["LABEL_LIST"],
             Literal["INVERTED"],
             Literal["FTS"],
+            Literal["NGRAM"],
         ],
         name: Optional[str] = None,
         *,
@@ -1547,6 +1548,10 @@ def create_scalar_index(
           contains lists of tags (e.g. ``["tag1", "tag2", "tag3"]``) can be indexed
           with a ``LABEL_LIST`` index.  This index can only speedup queries with
           ``array_has_any`` or ``array_has_all`` filters.
+        * ``NGRAM``. A special index that is used to index string columns.  This index
+          creates a bitmap for each ngram in the string.  By default we use trigrams.
+          This index can currently speed up queries using the ``contains`` function
+          in filters.
         * ``FTS/INVERTED``. It is used to index document columns. This index
           can conduct full-text searches. For example, a column that contains any word
           of query string "hello world". The results will be ranked by BM25.
@@ -1564,7 +1569,7 @@ def create_scalar_index(
             or string column.
         index_type : str
             The type of the index.  One of ``"BTREE"``, ``"BITMAP"``,
-            ``"LABEL_LIST"``, "FTS" or ``"INVERTED"``.
+            ``"LABEL_LIST"``, ``"NGRAM"``, ``"FTS"`` or ``"INVERTED"``.
         name : str, optional
             The index name. If not provided, it will be generated from the
             column name.
@@ -1651,10 +1656,10 @@ def create_scalar_index(
             raise KeyError(f"{column} not found in schema")
 
         index_type = index_type.upper()
-        if index_type not in ["BTREE", "BITMAP", "LABEL_LIST", "INVERTED"]:
+        if index_type not in ["BTREE", "BITMAP", "NGRAM", "LABEL_LIST", "INVERTED"]:
             raise NotImplementedError(
                 (
-                    'Only "BTREE", "LABEL_LIST", "INVERTED", '
+                    'Only "BTREE", "LABEL_LIST", "INVERTED", "NGRAM", '
                     'or "BITMAP" are supported for '
                     f"scalar columns.  Received {index_type}",
                 )
@@ -1676,6 +1681,9 @@ def create_scalar_index(
         elif index_type == "LABEL_LIST":
             if not pa.types.is_list(field.type):
                 raise TypeError(f"LABEL_LIST index column {column} must be a list")
+        elif index_type == "NGRAM":
+            if not pa.types.is_string(field.type):
+                raise TypeError(f"NGRAM index column {column} must be a string")
         elif index_type in ["INVERTED", "FTS"]:
             if not pa.types.is_string(field.type) and not pa.types.is_large_string(
                 field.type

python/python/tests/test_scalar_index.py

@@ -535,6 +535,43 @@ def test_bitmap_index(tmp_path: Path):
     assert indices[0]["type"] == "Bitmap"
 
 
+def test_ngram_index(tmp_path: Path):
+    """Test create ngram index"""
+    tbl = pa.Table.from_arrays(
+        [
+            pa.array(
+                [["apple", "apples", "banana", "coconut"][i % 4] for i in range(100)]
+            )
+        ],
+        names=["words"],
+    )
+    dataset = lance.write_dataset(tbl, tmp_path / "dataset")
+    dataset.create_scalar_index("words", index_type="NGRAM")
+    indices = dataset.list_indices()
+    assert len(indices) == 1
+    assert indices[0]["type"] == "NGram"
+
+    scan_plan = dataset.scanner(filter="contains(words, 'apple')").explain_plan(True)
+    assert "MaterializeIndex" in scan_plan
+
+    assert dataset.to_table(filter="contains(words, 'apple')").num_rows == 50
+    assert dataset.to_table(filter="contains(words, 'banana')").num_rows == 25
+    assert dataset.to_table(filter="contains(words, 'coconut')").num_rows == 25
+    assert dataset.to_table(filter="contains(words, 'apples')").num_rows == 25
+    assert (
+        dataset.to_table(
+            filter="contains(words, 'apple') AND contains(words, 'banana')"
+        ).num_rows
+        == 0
+    )
+    assert (
+        dataset.to_table(
+            filter="contains(words, 'apple') OR contains(words, 'banana')"
+        ).num_rows
+        == 75
+    )
+
+
 def test_null_handling(tmp_path: Path):
     tbl = pa.table(
         {
@@ -577,13 +614,15 @@ def test_scalar_index_with_nulls(tmp_path):
             "numeric_float": [0.1, None] * (test_table_size // 2),
             "boolean_col": [True, None] * (test_table_size // 2),
             "timestamp_col": [datetime(2023, 1, 1), None] * (test_table_size // 2),
+            "ngram_col": ["apple", None] * (test_table_size // 2),
         }
     )
     ds = lance.write_dataset(test_table, tmp_path)
     ds.create_scalar_index("inner_id", index_type="BTREE")
     ds.create_scalar_index("category", index_type="BTREE")
     ds.create_scalar_index("boolean_col", index_type="BTREE")
     ds.create_scalar_index("timestamp_col", index_type="BTREE")
+    ds.create_scalar_index("ngram_col", index_type="NGRAM")
     # Test querying with filters on columns with nulls.
     k = test_table_size // 2
     result = ds.to_table(filter="category = 'a'", limit=k)
@@ -594,6 +633,14 @@ def test_scalar_index_with_nulls(tmp_path):
     result = ds.to_table(filter="timestamp_col IS NOT NULL", limit=k)
     assert len(result) == k
 
+    # Ensure ngram index works with nulls
+    result = ds.to_table(filter="ngram_col = 'apple'")
+    assert len(result) == k
+    result = ds.to_table(filter="ngram_col IS NULL")
+    assert len(result) == k
+    result = ds.to_table(filter="contains(ngram_col, 'appl')")
+    assert len(result) == k
+
 
 def test_label_list_index(tmp_path: Path):
     tags = pa.array(["tag1", "tag2", "tag3", "tag4", "tag5", "tag6", "tag7"])
@@ -615,11 +662,12 @@ def test_create_index_empty_dataset(tmp_path: Path):
             pa.field("bitmap", pa.int32()),
             pa.field("label_list", pa.list_(pa.string())),
             pa.field("inverted", pa.string()),
+            pa.field("ngram", pa.string()),
         ]
     )
     ds = lance.write_dataset([], tmp_path, schema=schema)
 
-    for index_type in ["BTREE", "BITMAP", "LABEL_LIST", "INVERTED"]:
+    for index_type in ["BTREE", "BITMAP", "LABEL_LIST", "INVERTED", "NGRAM"]:
         ds.create_scalar_index(index_type.lower(), index_type=index_type)
 
     # Make sure the empty index doesn't cause searches to fail
@@ -630,6 +678,7 @@ def test_create_index_empty_dataset(tmp_path: Path):
                 "bitmap": pa.array([1], pa.int32()),
                 "label_list": [["foo", "bar"]],
                 "inverted": ["blah"],
+                "ngram": ["apple"],
             }
         )
     )
@@ -643,6 +692,9 @@ def test_searches():
         assert ds.to_table(filter="array_has_any(label_list, ['oof'])").num_rows == 0
         assert ds.to_table(filter="inverted = 'blah'").num_rows == 1
         assert ds.to_table(filter="inverted = 'halb'").num_rows == 0
+        assert ds.to_table(filter="contains(ngram, 'apple')").num_rows == 1
+        assert ds.to_table(filter="contains(ngram, 'banana')").num_rows == 0
+        assert ds.to_table(filter="ngram = 'apple'").num_rows == 1
 
     test_searches()
 
@@ -659,32 +711,47 @@ def test_searches():
 
 def test_optimize_no_new_data(tmp_path: Path):
     tbl = pa.table(
-        {"btree": pa.array([None], pa.int64()), "bitmap": pa.array([None], pa.int64())}
+        {
+            "btree": pa.array([None], pa.int64()),
+            "bitmap": pa.array([None], pa.int64()),
+            "ngram": pa.array([None], pa.string()),
+        }
     )
     dataset = lance.write_dataset(tbl, tmp_path)
     dataset.create_scalar_index("btree", index_type="BTREE")
     dataset.create_scalar_index("bitmap", index_type="BITMAP")
+    dataset.create_scalar_index("ngram", index_type="NGRAM")
 
     assert dataset.to_table(filter="btree IS NULL").num_rows == 1
     assert dataset.to_table(filter="bitmap IS NULL").num_rows == 1
+    assert dataset.to_table(filter="ngram IS NULL").num_rows == 1
 
     dataset.insert([], schema=tbl.schema)
     dataset.optimize.optimize_indices()
 
     assert dataset.to_table(filter="btree IS NULL").num_rows == 1
     assert dataset.to_table(filter="bitmap IS NULL").num_rows == 1
+    assert dataset.to_table(filter="ngram IS NULL").num_rows == 1
 
     dataset.insert(pa.table({"btree": [2]}))
     dataset.optimize.optimize_indices()
 
     assert dataset.to_table(filter="btree IS NULL").num_rows == 1
     assert dataset.to_table(filter="bitmap IS NULL").num_rows == 2
+    assert dataset.to_table(filter="ngram IS NULL").num_rows == 2
 
     dataset.insert(pa.table({"bitmap": [2]}))
     dataset.optimize.optimize_indices()
 
     assert dataset.to_table(filter="btree IS NULL").num_rows == 2
     assert dataset.to_table(filter="bitmap IS NULL").num_rows == 2
+    assert dataset.to_table(filter="ngram IS NULL").num_rows == 3
+
+    dataset.insert(pa.table({"ngram": ["apple"]}))
+
+    assert dataset.to_table(filter="btree IS NULL").num_rows == 3
+    assert dataset.to_table(filter="bitmap IS NULL").num_rows == 3
+    assert dataset.to_table(filter="ngram IS NULL").num_rows == 3
 
 
 def test_drop_index(tmp_path):
@@ -694,14 +761,16 @@ def test_drop_index(tmp_path):
             "btree": list(range(test_table_size)),
             "bitmap": list(range(test_table_size)),
             "fts": ["a" for _ in range(test_table_size)],
+            "ngram": ["a" for _ in range(test_table_size)],
         }
     )
     ds = lance.write_dataset(test_table, tmp_path)
     ds.create_scalar_index("btree", index_type="BTREE")
     ds.create_scalar_index("bitmap", index_type="BITMAP")
     ds.create_scalar_index("fts", index_type="INVERTED")
+    ds.create_scalar_index("ngram", index_type="NGRAM")
 
-    assert len(ds.list_indices()) == 3
+    assert len(ds.list_indices()) == 4
 
     # Attempt to drop index (name does not exist)
     with pytest.raises(RuntimeError, match="index not found"):
@@ -717,3 +786,4 @@ def test_drop_index(tmp_path):
     assert ds.to_table(filter="btree = 1").num_rows == 1
     assert ds.to_table(filter="bitmap = 1").num_rows == 1
     assert ds.to_table(filter="fts = 'a'").num_rows == test_table_size
+    assert ds.to_table(filter="contains(ngram, 'a')").num_rows == test_table_size

python/src/dataset.rs

@@ -1176,6 +1176,7 @@ impl Dataset {
         let idx_type = match index_type.as_str() {
             "BTREE" => IndexType::Scalar,
             "BITMAP" => IndexType::Bitmap,
+            "NGRAM" => IndexType::NGram,
             "LABEL_LIST" => IndexType::LabelList,
             "INVERTED" | "FTS" => IndexType::Inverted,
             "IVF_FLAT" | "IVF_PQ" | "IVF_HNSW_PQ" | "IVF_HNSW_SQ" => IndexType::Vector,
@@ -1193,6 +1194,9 @@ impl Dataset {
                 // Temporary workaround until we add support for auto-detection of scalar index type
                 force_index_type: Some(ScalarIndexType::Bitmap),
             }),
+            "NGRAM" => Box::new(ScalarIndexParams {
+                force_index_type: Some(ScalarIndexType::NGram),
+            }),
             "LABEL_LIST" => Box::new(ScalarIndexParams {
                 force_index_type: Some(ScalarIndexType::LabelList),
             }),

rust/lance-core/src/datatypes/schema.rs

@@ -828,6 +828,16 @@ impl Projection {
         }
     }
 
+    pub fn with_row_id(mut self) -> Self {
+        self.with_row_id = true;
+        self
+    }
+
+    pub fn with_row_addr(mut self) -> Self {
+        self.with_row_addr = true;
+        self
+    }
+
     /// Add a column (and any of its parents) to the projection from a string reference
     pub fn union_column(mut self, column: impl AsRef<str>, on_missing: OnMissing) -> Result<Self> {
         let column = column.as_ref();
@@ -855,6 +865,11 @@ impl Projection {
         self.field_ids.contains(&id)
     }
 
+    /// True if the projection selects fields other than the row id / addr
+    pub fn has_data_fields(&self) -> bool {
+        !self.field_ids.is_empty()
+    }
+
     /// Add multiple columns (and their parents) to the projection
     pub fn union_columns(
         mut self,

rust/lance-core/src/utils/mask.rs

@@ -10,7 +10,7 @@ use arrow_array::{Array, BinaryArray, GenericBinaryArray};
 use arrow_buffer::{Buffer, NullBuffer, OffsetBuffer};
 use byteorder::{ReadBytesExt, WriteBytesExt};
 use deepsize::DeepSizeOf;
-use roaring::{MultiOps, RoaringBitmap};
+use roaring::{MultiOps, RoaringBitmap, RoaringTreemap};
 
 use crate::Result;
 
@@ -706,6 +706,16 @@ impl<'a> FromIterator<&'a u64> for RowIdTreeMap {
     }
 }
 
+impl From<RoaringTreemap> for RowIdTreeMap {
+    fn from(roaring: RoaringTreemap) -> Self {
+        let mut inner = BTreeMap::new();
+        for (fragment, set) in roaring.bitmaps() {
+            inner.insert(fragment, RowIdSelection::Partial(set.clone()));
+        }
+        Self { inner }
+    }
+}
+
 impl Extend<u64> for RowIdTreeMap {
     fn extend<T: IntoIterator<Item = u64>>(&mut self, iter: T) {
         for row_id in iter {

rust/lance-index/Cargo.toml

@@ -113,6 +113,10 @@ harness = false
 name = "sq"
 harness = false
 
+[[bench]]
+name = "ngram"
+harness = false
+
 [[bench]]
 name = "inverted"
 harness = false