ann-filtering-benchmark directory

.gitignore

@@ -26,3 +26,5 @@ sqlite-vec.h
 tmp/
 
 poetry.lock
+
+*.jsonl

sqlite-vec.c

@@ -5972,51 +5972,52 @@ int vec0_set_metadata_filter_bitmap(
           }
           break;
         }
-        case VEC0_METADATA_OPERATOR_GT: {
+        case VEC0_METADATA_OPERATOR_NE: {
           for(int i = 0; i < size; i++) {
             u8 * view = &((u8*) buffer)[i * VEC0_METADATA_TEXT_VIEW_BUFFER_LENGTH];
             int n = ((int*) view)[0];
             char * s = (char *) &view[4];
-            bitmap_set(b, i, strncmp(s, target, n) > 0);
+            bitmap_set(b, i, strncmp(s, target, n) != 0);
           }
           break;
         }
-        case VEC0_METADATA_OPERATOR_LE: {
+        case VEC0_METADATA_OPERATOR_GT: {
           for(int i = 0; i < size; i++) {
             u8 * view = &((u8*) buffer)[i * VEC0_METADATA_TEXT_VIEW_BUFFER_LENGTH];
             int n = ((int*) view)[0];
             char * s = (char *) &view[4];
-            bitmap_set(b, i, strncmp(s, target, n) <= 0);
+            bitmap_set(b, i, strncmp(s, target, n) > 0);
           }
           break;
         }
-        case VEC0_METADATA_OPERATOR_LT: {
+        case VEC0_METADATA_OPERATOR_GE: {
           for(int i = 0; i < size; i++) {
             u8 * view = &((u8*) buffer)[i * VEC0_METADATA_TEXT_VIEW_BUFFER_LENGTH];
             int n = ((int*) view)[0];
             char * s = (char *) &view[4];
-            bitmap_set(b, i, strncmp(s, target, n) < 0);
+            bitmap_set(b, i, strncmp(s, target, n) >= 0);
           }
           break;
         }
-        case VEC0_METADATA_OPERATOR_GE: {
+        case VEC0_METADATA_OPERATOR_LE: {
           for(int i = 0; i < size; i++) {
             u8 * view = &((u8*) buffer)[i * VEC0_METADATA_TEXT_VIEW_BUFFER_LENGTH];
             int n = ((int*) view)[0];
             char * s = (char *) &view[4];
-            bitmap_set(b, i, strncmp(s, target, n) >= 0);
+            bitmap_set(b, i, strncmp(s, target, n) <= 0);
           }
           break;
         }
-        case VEC0_METADATA_OPERATOR_NE: {
+        case VEC0_METADATA_OPERATOR_LT: {
           for(int i = 0; i < size; i++) {
             u8 * view = &((u8*) buffer)[i * VEC0_METADATA_TEXT_VIEW_BUFFER_LENGTH];
             int n = ((int*) view)[0];
             char * s = (char *) &view[4];
-            bitmap_set(b, i, strncmp(s, target, n) != 0);
+            bitmap_set(b, i, strncmp(s, target, n) < 0);
           }
           break;
         }
+
       }
       break;
     }

tests/afbd/.gitignore

@@ -0,0 +1 @@
+*.tgz

tests/afbd/.python-version

@@ -0,0 +1 @@
+3.12

tests/afbd/Makefile

@@ -0,0 +1,9 @@
+random_ints_1m.tgz:
+	curl -o $@ https://storage.googleapis.com/ann-filtered-benchmark/datasets/random_ints_1m.tgz
+
+random_float_1m.tgz:
+	curl -o $@ https://storage.googleapis.com/ann-filtered-benchmark/datasets/random_float_1m.tgz
+
+random_keywords_1m.tgz:
+	curl -o $@ https://storage.googleapis.com/ann-filtered-benchmark/datasets/random_keywords_1m.tgz
+all: random_ints_1m.tgz random_float_1m.tgz random_keywords_1m.tgz

tests/afbd/README.md

@@ -0,0 +1,12 @@
+
+# hnm
+
+```
+tar -xOzf hnm.tgz ./tests.jsonl  > tests.jsonl
+solite q "select group_concat(distinct key) from lines_read('tests.jsonl'), json_each(line -> '$.conditions.and[0]')"
+```
+
+
+```
+> python test-afbd.py build hnm.tgz --metadata product_group_name,colour_group_name,index_group_name,perceived_colour_value_name,section_name,product_type_name,department_name,graphical_appearance_name,garment_group_name,perceived_colour_master_name
+```

tests/afbd/test-afbd.py

@@ -0,0 +1,215 @@
+import numpy as np
+from tqdm import tqdm
+from deepdiff import DeepDiff
+
+import tarfile
+import json
+from io import BytesIO
+import sqlite3
+from typing import List
+from struct import pack
+import time
+from pathlib import Path
+import argparse
+
+
+def serialize_float32(vector: List[float]) -> bytes:
+    """Serializes a list of floats into the "raw bytes" format sqlite-vec expects"""
+    return pack("%sf" % len(vector), *vector)
+
+
+def build_command(file_path, metadata_set=None):
+    if metadata_set:
+        metadata_set = set(metadata_set.split(","))
+
+    file_path = Path(file_path)
+    print(f"reading {file_path}...")
+    t0 = time.time()
+    with tarfile.open(file_path, "r:gz") as archive:
+        for file in archive:
+            if file.name == "./payloads.jsonl":
+                payloads = [
+                    json.loads(line)
+                    for line in archive.extractfile(file.name).readlines()
+                ]
+            if file.name == "./tests.jsonl":
+                tests = [
+                    json.loads(line)
+                    for line in archive.extractfile(file.name).readlines()
+                ]
+            if file.name == "./vectors.npy":
+                f = BytesIO()
+                f.write(archive.extractfile(file.name).read())
+                f.seek(0)
+                vectors = np.load(f)
+
+    assert payloads is not None
+    assert tests is not None
+    assert vectors is not None
+    dimensions = vectors.shape[1]
+    metadata_columns = sorted(list(payloads[0].keys()))
+
+    def col_type(v):
+        if isinstance(v, int):
+            return "integer"
+        if isinstance(v, float):
+            return "float"
+        if isinstance(v, str):
+            return "text"
+        raise Exception(f"Unknown column type: {v}")
+
+    metadata_columns_types = [col_type(payloads[0][col]) for col in metadata_columns]
+
+    print(time.time() - t0)
+    t0 = time.time()
+    print("seeding...")
+
+    db = sqlite3.connect(f"{file_path.stem}.db")
+    db.execute("PRAGMA page_size = 16384")
+    db.row_factory = sqlite3.Row
+    db.enable_load_extension(True)
+    db.load_extension("../../dist/vec0")
+    db.enable_load_extension(False)
+
+    with db:
+        db.execute("create table tests(data)")
+
+        for test in tests:
+            db.execute("insert into tests values (?)", [json.dumps(test)])
+
+    with db:
+        create_sql = f"create virtual table v using vec0(vector float[{dimensions}] distance_metric=cosine"
+        insert_sql = "insert into v(rowid, vector"
+        for name, type in zip(metadata_columns, metadata_columns_types):
+            if metadata_set:
+                if name in metadata_set:
+                    create_sql += f", {name} {type}"
+                else:
+                    create_sql += f", +{name} {type}"
+            else:
+                create_sql += f", {name} {type}"
+
+            insert_sql += f", {name}"
+        create_sql += ")"
+        insert_sql += ") values (" + ",".join("?" * (2 + len(metadata_columns))) + ")"
+        print(create_sql)
+        print(insert_sql)
+
+        db.execute(create_sql)
+
+        for idx, (payload, vector) in enumerate(
+            tqdm(zip(payloads, vectors), total=len(payloads))
+        ):
+            params = [idx, vector]
+            for c in metadata_columns:
+                params.append(payload[c])
+            db.execute(insert_sql, params)
+
+    print(time.time() - t0)
+
+
+def tests_command(file_path):
+    file_path = Path(file_path)
+    db = sqlite3.connect(f"{file_path.stem}.db")
+    db.execute("PRAGMA cache_size = -100000000")
+    db.row_factory = sqlite3.Row
+    db.enable_load_extension(True)
+    db.load_extension("../../dist/vec0")
+    db.enable_load_extension(False)
+
+    tests = [
+        json.loads(row["data"])
+        for row in db.execute("select data from tests limit 2000").fetchall()
+    ]
+
+    num_or_skips = 0
+    num_1off_errors = 0
+
+    t0 = time.time()
+    print("testing...")
+    for idx, test in enumerate(tqdm(tests)):
+        query = test["query"]
+        conditions = test["conditions"]
+        expected_closest_ids = test["closest_ids"]
+        expected_closest_scores = test["closest_scores"]
+        if "or" in conditions:
+            num_or_skips += 1
+            continue
+
+        sql = "select rowid, 1 - distance as similarity from v where vector match ? and k = ?"
+        params = [serialize_float32(query), len(expected_closest_ids)]
+
+        for condition in conditions["and"]:
+            assert len(condition.keys()) == 1
+            column = list(condition.keys())[0]
+            assert len(list(condition[column].keys())) == 1
+            condition_type = list(condition[column].keys())[0]
+            if condition_type == "match":
+                value = condition[column]["match"]["value"]
+                sql += f" and {column} = ?"
+                params.append(value)
+            elif condition_type == "range":
+                sql += f" and {column} between ? and ?"
+                params.append(condition[column]["range"]["gt"])
+                params.append(condition[column]["range"]["lt"])
+            else:
+                raise Exception(f"Unknown condition type: {condition_type}")
+
+        rows = db.execute(sql, params).fetchall()
+        actual_closest_ids = [row["rowid"] for row in rows]
+        matches = expected_closest_ids == actual_closest_ids
+        if not matches:
+            diff = DeepDiff(
+                expected_closest_ids, actual_closest_ids, ignore_order=False
+            )
+            assert len(list(diff.keys())) == 1
+            assert "values_changed" in diff.keys()
+            keys_changed = list(diff["values_changed"].keys())
+            if len(keys_changed) == 2:
+                akey, bkey = keys_changed
+                a = int(akey.lstrip("root[").rstrip("]"))
+                b = int(bkey.lstrip("root[").rstrip("]"))
+                assert abs(a - b) == 1
+                assert (
+                    diff["values_changed"][akey]["new_value"]
+                    == diff["values_changed"][bkey]["old_value"]
+                )
+                assert (
+                    diff["values_changed"][akey]["old_value"]
+                    == diff["values_changed"][bkey]["new_value"]
+                )
+            elif len(keys_changed) == 1:
+                v = int(akey.lstrip("root[").rstrip("]"))
+                assert v == len(expected_closest_ids)
+            else:
+                raise Exception("fuck")
+            num_1off_errors += 1
+        # print(closest_scores)
+        # print([row["similarity"] for row in rows])
+        # assert closest_scores == [row["similarity"] for row in rows]
+    print("Number skipped: ", num_or_skips)
+    print("Num 1 off errors: ", num_1off_errors)
+    print("1 off error rate: ", num_1off_errors / (len(tests) - num_or_skips))
+    print(time.time() - t0)
+    print("done")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="CLI tool")
+    subparsers = parser.add_subparsers(dest="command", required=True)
+
+    build_parser = subparsers.add_parser("build")
+    build_parser.add_argument("file", type=str, help="Path to input file")
+    build_parser.add_argument("--metadata", type=str, help="Metadata columns")
+    build_parser.set_defaults(func=lambda args: build_command(args.file, args.metadata))
+
+    tests_parser = subparsers.add_parser("test")
+    tests_parser.add_argument("file", type=str, help="Path to input file")
+    tests_parser.set_defaults(func=lambda args: tests_command(args.file))
+
+    args = parser.parse_args()
+    args.func(args)
+
+
+if __name__ == "__main__":
+    main()