@an5/adapters 0.2.11 → 0.2.14

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@an5/adapters",
-  "version": "0.2.11",
-  "description": "Standalone language adapters for AN5 ORM - TypeScript, Python, .NET, Go, Rust, and Google Sheets query execution.",
+  "version": "0.2.14",
+  "description": "Standalone language adapters for AN5 ORM - TypeScript, Python, .NET, Go, Rust, Java, Kotlin, Swift, and Google Sheets query execution.",
   "license": "MIT",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -23,6 +23,9 @@
     "./dotnet": "./dotnet/An5Adapter.cs",
     "./golang": "./golang/an5_adapter.go",
     "./rust": "./rust/Cargo.toml",
+    "./java": "./java/src/main/java",
+    "./kotlin": "./kotlin/src/main/kotlin",
+    "./swift": "./swift/Sources/An5Adapters",
     "./nbase": "./dist/nbase/index.js"
   },
   "author": "an5ORM",
@@ -40,31 +43,54 @@
     "dotnet",
     "go",
     "golang",
-    "rust"
+    "rust",
+    "java",
+    "kotlin",
+    "swift"
   ],
   "files": [
     "dist/**/*.{js,d.ts}",
+    "native/sqlite/**/*",
+    "!native/sqlite/build/**/*",
+    "scripts/sqlite-native-build.js",
     "typescript/src/**/*.ts",
+    "typescript/src/nbase/**/*.ts",
     "python/**/*.py",
     "dotnet/**/*",
     "golang/**/*",
     "rust/Cargo.toml",
     "rust/src/**/*",
-    "README.md",
-    "typescript/src/nbase/**/*.ts"
+    "java/pom.xml",
+    "java/src/main/java/**/*",
+    "java/src/test/java/**/*",
+    "kotlin/build.gradle.kts",
+    "kotlin/settings.gradle.kts",
+    "kotlin/src/main/kotlin/**/*",
+    "kotlin/src/test/kotlin/**/*",
+    "swift/Package.swift",
+    "swift/Sources/**/*",
+    "swift/Tests/**/*",
+    "README.md"
   ],
   "scripts": {
     "build": "tsc -p tsconfig.json",
     "pretest": "npm run build",
     "prepack": "npm run build",
     "build:python": "(python -m build --outdir dist-py || python3 -m build --outdir dist-py)",
-    "test": "node test/unit.test.js && node test/package.test.js && node test/nbase.test.js && node --test test/version-sync.test.js && node --test test/query-semantics.test.js && node test/query-relations.integration.test.js && node --test test/sheets-oauth.test.js",
+    "test": "node test/unit.test.js && node test/package.test.js && node test/nbase.test.js && node test/sqlite-vector.test.js && node --test test/version-sync.test.js && node --test test/query-semantics.test.js && node test/query-relations.integration.test.js && node --test test/sheets-oauth.test.js",
     "test:package:smoke": "node scripts/package-smoke.js",
     "test:integration:live": "node test/live-db.integration.test.js && node test/query-relations.integration.test.js",
     "test:python": "node scripts/python-tests.js",
     "test:dotnet": "node scripts/dotnet-compile-check.js && node scripts/dotnet-sqlite-smoke.js",
     "test:go": "cd golang && go test ./...",
-    "test:rust": "node scripts/rust-compile-check.js"
+    "test:rust": "node scripts/rust-compile-check.js",
+    "test:java": "node scripts/java-compile-check.js && node scripts/java-sqlite-smoke.js",
+    "test:kotlin": "node scripts/kotlin-compile-check.js",
+    "test:gradle": "node scripts/kotlin-gradle-build.js",
+    "test:swift": "node scripts/swift-test.js",
+    "build:sqlite:native": "node scripts/sqlite-native-build.js",
+    "test:sqlite:native": "node scripts/sqlite-native-test.js",
+    "bench:sqlite:native": "node scripts/sqlite-native-test.js --benchmark"
   },
   "dependencies": {
     "googleapis": "^183.0.0",

package/python/an5_adapter.py

@@ -4,15 +4,17 @@ from typing import Dict, List, Optional, Any
 
 try:
     from .base import DIALECT_MSSQL, DIALECT_POSTGRES, DIALECT_SQLITE, detect_dialect, set_adapter_metadata
+    from .base.vectors import decode_vector, encode_vector, vector_distance
     from .mssql import connect as connect_mssql
     from .postgres import connect as connect_postgres
-    from .sqlite import connect as connect_sqlite, is_memory as is_sqlite_memory
+    from .sqlite import connect as connect_sqlite, is_memory as is_sqlite_memory, register_vector_functions as register_sqlite_vector_functions
     from .table_client import AdapterTableClient, ViewClient
 except ImportError:
     from base import DIALECT_MSSQL, DIALECT_POSTGRES, DIALECT_SQLITE, detect_dialect, set_adapter_metadata
+    from base.vectors import decode_vector, encode_vector, vector_distance
     from mssql import connect as connect_mssql
     from postgres import connect as connect_postgres
-    from sqlite import connect as connect_sqlite, is_memory as is_sqlite_memory
+    from sqlite import connect as connect_sqlite, is_memory as is_sqlite_memory, register_vector_functions as register_sqlite_vector_functions
     from table_client import AdapterTableClient, ViewClient
 
 # Backward-compatible aliases used by tests and older imports.
@@ -23,9 +25,15 @@ except ImportError:
     from mssql import parse_connection_string as _parse_connection_string
 
 class An5Adapter:
-    def __init__(self, connection_string: str):
+    def __init__(self, connection_string: str, sqlite_vec: Optional[str] = None, vector_strategy: Optional[str] = None):
         self._dialect = detect_dialect(connection_string)
         self._conn_str = connection_string
+        # SQLite only. `sqlite_vec` names the extension binary to load;
+        # `vector_strategy` pins one of "sqlite-vec" | "udf" | "sql" | "memory"
+        # instead of probing for the fastest one available.
+        self._sqlite_vec = sqlite_vec
+        self.vector_strategy = vector_strategy
+        self._vector_supports: Dict[int, Any] = {}
         # The open connection of `transaction()`, when there is one. Every
         # statement in the callback has to use this very connection, otherwise a
         # rollback returns to a different, already committed connection — that
@@ -39,9 +47,29 @@ class An5Adapter:
         if self._dialect == DIALECT_POSTGRES:
             return connect_postgres(self._conn_str)
         if self._dialect == DIALECT_SQLITE:
-            return connect_sqlite(self._conn_str)
+            conn = connect_sqlite(self._conn_str)
+            support = register_sqlite_vector_functions(conn, self._sqlite_vec)
+            support.capabilities()
+            self._sqlite_vector_support = support
+            return conn
         return connect_mssql(self._conn_str)
 
+    def vector_support(self):
+        """The SQLite vector hooks of the current connection, or None.
+
+        They register the distance functions SQLite ranks with and load the
+        sqlite-vec extension. A connection with none still searches, by falling
+        back to `json_each` or to Python.
+        """
+        if self._dialect != DIALECT_SQLITE:
+            return None
+        conn, owned = self._acquire()
+        try:
+            return self._sqlite_vector_support
+        finally:
+            if owned:
+                conn.close()
+
     def _acquire(self):
         """Returns (connection, whether we opened it and must close it)."""
         if self._txn_conn is not None:
@@ -167,11 +195,14 @@ class An5Adapter:
             if owned:
                 conn.close()
 
-def create_an5_adapter(connection_string: str) -> An5Adapter:
-    return An5Adapter(connection_string)
+def create_an5_adapter(connection_string: str, sqlite_vec: Optional[str] = None, vector_strategy: Optional[str] = None) -> An5Adapter:
+    return An5Adapter(connection_string, sqlite_vec=sqlite_vec, vector_strategy=vector_strategy)
 
 __all__ = [
     "An5Adapter",
+    "encode_vector",
+    "decode_vector",
+    "vector_distance",
     "AdapterTableClient",
     "ViewClient",
     "create_an5_adapter",

package/python/base/__init__.py

@@ -1,3 +1,18 @@
 from .dialects import DIALECT_MSSQL, DIALECT_POSTGRES, DIALECT_SQLITE, detect_dialect
 from .metadata import model_to_table, model_fields, relation_map, set_adapter_metadata, get_relations_for_model, get_model_to_table, get_fields_for_model
 from .sql import _quote, _parse_where, _build_order_by, _resolve_table, _quote_table
+from .vectors import (
+    AN5_VECTOR_FUNCTIONS,
+    SQLITE_VEC_FUNCTIONS,
+    SqliteVectorSupport,
+    decode_vector,
+    encode_vector,
+    is_vector_field,
+    make_distance_function,
+    needs_vector_encoding,
+    parse_vector,
+    plan_sqlite_vector_strategies,
+    read_sqlite_column_type,
+    run_sqlite_vector_search,
+    vector_distance,
+)

package/python/base/vectors.py

@@ -0,0 +1,431 @@
+"""SQLite vector search — codec, distance functions and the strategy plan.
+
+SQLite has no vector type, so a `VECTOR(n)` column stores a BLOB of little-endian
+float32 values (see `encode_vector`) and is ranked in one of four ways, in this
+order:
+
+   1. ``sqlite-vec``  native scalar distance functions, when the extension loads.
+   2. ``udf``         ``an5_vec_cosine`` / ``an5_vec_l2`` / ``an5_vec_ip``
+      registered with the driver. Reads the BLOB and the legacy JSON text.
+   3. ``sql``         ``json_each`` brute force in plain SQL. Needs no user
+      function, but only reaches rows stored as JSON text.
+   4. ``memory``      the column is loaded and scored in Python.
+
+Strategies 1-3 rank inside the database and transfer only the matching rows,
+which is why they are preferred over the in-memory path.
+"""
+
+import json
+import math
+import struct
+from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple
+
+METRIC_COSINE = "cosine"
+METRIC_EUCLIDEAN = "euclidean"
+METRIC_DOT = "dot"
+
+# Scalar functions the adapter registers itself.
+AN5_VECTOR_FUNCTIONS: Dict[str, str] = {
+    METRIC_COSINE: "an5_vec_cosine",
+    METRIC_EUCLIDEAN: "an5_vec_l2",
+    METRIC_DOT: "an5_vec_ip",
+}
+
+# Scalar functions shipped by the sqlite-vec extension.
+SQLITE_VEC_FUNCTIONS: Dict[str, str] = {
+    METRIC_COSINE: "vec_distance_cosine",
+    METRIC_EUCLIDEAN: "vec_distance_l2",
+    METRIC_DOT: "vec_distance_ip",
+}
+
+STRATEGY_SQLITE_VEC = "sqlite-vec"
+STRATEGY_UDF = "udf"
+STRATEGY_SQL = "sql"
+STRATEGY_MEMORY = "memory"
+
+_BYTES_PER_FLOAT = 4
+_FLOAT32 = struct.Struct("<f")
+
+
+def normalize_metric(metric: Optional[str]) -> str:
+    return metric if metric in (METRIC_EUCLIDEAN, METRIC_DOT) else METRIC_COSINE
+
+
+# ─── Codec ───────────────────────────────────────────────────────────────────────
+
+
+def parse_vector(value: Any) -> Optional[List[float]]:
+    """Reads a numeric sequence, or JSON text holding one."""
+    if value is None or isinstance(value, (str, bytes, bytearray, memoryview)):
+        if isinstance(value, (bytes, bytearray, memoryview)):
+            return None
+        text = str(value).strip()
+        if not text.startswith("["):
+            return None
+        try:
+            return parse_vector(json.loads(text))
+        except Exception:
+            return None
+    if isinstance(value, dict):
+        return None
+    try:
+        values = [float(v) for v in value]
+    except (TypeError, ValueError):
+        return None
+    if not values or any(math.isnan(v) or math.isinf(v) for v in values):
+        return None
+    return values
+
+
+def decode_vector(value: Any) -> Optional[List[float]]:
+    """Decodes a stored vector: float32 BLOB, legacy JSON text, or a sequence."""
+    direct = parse_vector(value)
+    if direct:
+        return direct
+    if isinstance(value, (bytes, bytearray, memoryview)):
+        raw = bytes(value)
+        if not raw or len(raw) % _BYTES_PER_FLOAT != 0:
+            return None
+        count = len(raw) // _BYTES_PER_FLOAT
+        return list(_FLOAT32.unpack_from(raw, i * _BYTES_PER_FLOAT)[0] for i in range(count))
+    return None
+
+
+def encode_vector(values: Sequence[float]) -> bytes:
+    """Encodes a vector as the little-endian float32 BLOB the column stores."""
+    return b"".join(_FLOAT32.pack(float(v)) for v in values)
+
+
+def needs_vector_encoding(value: Any) -> bool:
+    """True when the value still has to become bytes on its way into a column."""
+    if isinstance(value, (bytes, bytearray, memoryview, str)):
+        return False
+    if value is None:
+        return False
+    return isinstance(value, (list, tuple)) or (
+        hasattr(value, "__iter__") and not isinstance(value, (dict, int, float, bool))
+    )
+
+
+def is_vector_field(definition: Any) -> bool:
+    """True when generated metadata describes a `VECTOR(n)` column."""
+    if not definition:
+        return False
+    if isinstance(definition, str):
+        return "[]" in definition or definition.strip().upper().startswith("VECTOR")
+    if not isinstance(definition, dict):
+        return False
+    if definition.get("kind") == "vector":
+        return True
+    sql = str(definition.get("sql") or definition.get("type") or "")
+    if sql.strip().upper().startswith("VECTOR") and sql.strip()[6:7] in ("", "("):
+        return True
+    return "[]" in str(definition.get("ts") or "")
+
+
+# ─── Distance ─────────────────────────────────────────────────────────────────────
+
+
+def vector_distance(a: Sequence[float], b: Sequence[float], metric: str) -> Optional[float]:
+    """Distance between two equal-length vectors, lower is closer.
+
+    Mirrors the SQL functions so a search ranks the same way whichever strategy
+    ran.
+    """
+    if a is None or b is None or len(a) != len(b) or len(a) == 0:
+        return None
+    resolved = normalize_metric(metric)
+    if resolved == METRIC_EUCLIDEAN:
+        return math.sqrt(sum((float(x) - float(y)) ** 2 for x, y in zip(a, b)))
+    dot = sum(float(x) * float(y) for x, y in zip(a, b))
+    if resolved == METRIC_DOT:
+        return -dot
+    m1 = math.sqrt(sum(float(x) ** 2 for x in a))
+    m2 = math.sqrt(sum(float(y) ** 2 for y in b))
+    if not m1 or not m2:
+        return 1.0
+    return 1.0 - dot / (m1 * m2)
+
+
+def make_distance_function(metric: str) -> Callable[[Any, Any], Optional[float]]:
+    """The scalar function registered with the driver for one metric."""
+
+    def distance(left: Any, right: Any) -> Optional[float]:
+        a = decode_vector(left)
+        b = decode_vector(right)
+        if not a or not b:
+            return None
+        return vector_distance(a, b, metric)
+
+    return distance
+
+
+# ─── Query building ───────────────────────────────────────────────────────────────
+
+
+def _json_text(column: str) -> str:
+    """`json_valid` fails on a float32 BLOB, so the JSON path needs this guard."""
+    return f"CASE WHEN json_valid({column}) THEN {column} ELSE '[]' END"
+
+
+def _json_distance_expr(metric: str, column: str) -> str:
+    row = _json_text(column)
+    same_length = f"(SELECT COUNT(*) FROM json_each({row})) = (SELECT COUNT(*) FROM q)"
+    terms = {
+        METRIC_COSINE: (
+            f"1.0 - (SELECT SUM(je.value * q.v) FROM json_each({row}) je JOIN q ON q.k = je.key)"
+            f" / NULLIF(sqrt((SELECT SUM(q.v * q.v) FROM q))"
+            f" * sqrt((SELECT SUM(je.value * je.value) FROM json_each({row}) je)), 0)"
+        ),
+        METRIC_EUCLIDEAN: (
+            f"sqrt((SELECT SUM((je.value - q.v) * (je.value - q.v))"
+            f" FROM json_each({row}) je JOIN q ON q.k = je.key))"
+        ),
+        METRIC_DOT: f"-(SELECT SUM(je.value * q.v) FROM json_each({row}) je JOIN q ON q.k = je.key)",
+    }
+    return f"CASE WHEN {same_length} THEN {terms[metric]} END"
+
+
+def build_sqlite_vector_query(
+    strategy: str,
+    metric: str,
+    table: str,
+    column: str,
+    vector: Sequence[float],
+    take: int,
+    tail: str = "",
+    params: Optional[List] = None,
+) -> Tuple[str, List]:
+    """Ranking SQL for one strategy, with the query vector appended to `params`.
+
+    `sqlite-vec` and `udf` bind the query vector as a float32 BLOB; `sql` binds
+    it as JSON text because `json_each` reads text.
+    """
+    bound = []
+    metric = normalize_metric(metric)
+
+    def rank(inner: str) -> str:
+        # A row whose stored vector cannot be scored (a NULL, a text column,
+        # another dimension) yields a NULL distance; those rows are dropped
+        # rather than reported, so a caller never has to sort them out.
+        return (f"SELECT * FROM ({inner}) AS an5_ranked"
+                f" WHERE distance IS NOT NULL ORDER BY distance ASC LIMIT {int(take)}")
+
+
+    if strategy == STRATEGY_SQLITE_VEC:
+        bound.append(encode_vector(vector))
+        distance = f"CASE WHEN vec_length(vec_f32({column})) = {len(vector)} THEN {SQLITE_VEC_FUNCTIONS[metric]}(vec_f32({column}), ?) END"
+        # vec_f32 accepts both JSON text and float32 BLOBs.
+        query = rank(
+            f"SELECT *, {distance} AS distance FROM {table}"
+            + (tail + " AND " if tail else " WHERE ")
+            + f"{column} IS NOT NULL"
+        )
+        return query, bound + list(params or [])
+
+    if strategy == STRATEGY_UDF:
+        bound.append(encode_vector(vector))
+        distance = f"{AN5_VECTOR_FUNCTIONS[metric]}({column}, ?)"
+        return rank(f"SELECT *, {distance} AS distance FROM {table}{tail}"), bound + list(params or [])
+
+    bound.append(json.dumps(list(vector)))
+    query = (
+        "WITH q AS (SELECT je.key AS k, CAST(je.value AS REAL) AS v FROM json_each(?) je) "
+        + rank(f"SELECT *, {_json_distance_expr(metric, column)} AS distance FROM {table}{tail}")
+    )
+    return query, bound + list(params or [])
+
+
+# ─── Capabilities ──────────────────────────────────────────────────────────────────
+
+
+def plan_sqlite_vector_strategies(
+    capabilities: Dict[str, bool],
+    declared_type: Optional[str],
+    preference: Optional[str] = None,
+) -> List[str]:
+    """The ordered strategies to try for this connection.
+
+    `declared_type` is the column's DDL type: the JSON strategy only reaches rows
+    stored as text, so a `BLOB` column drops it from the order.
+    """
+    if preference and preference != "auto":
+        # `memory` is the caller's own path, not a query, so it yields no plan.
+        return [] if preference == STRATEGY_MEMORY else [preference]
+    available: List[str] = []
+    if capabilities.get("vec"):
+        available.append(STRATEGY_SQLITE_VEC)
+    if capabilities.get("udf"):
+        available.append(STRATEGY_UDF)
+    if capabilities.get("json1"):
+        available.append(STRATEGY_SQL)
+    declared = str(declared_type or "").upper()
+    if declared and not any(token in declared for token in ("TEXT", "CHAR", "CLOB", "STRING", "JSON")):
+        # A BLOB column has no JSON to read, so json_each can only produce NULLs.
+        return [s for s in available if s != STRATEGY_SQL]
+    return available
+
+
+def read_sqlite_column_type(
+    exec_fn: Callable[[str, List], List[Dict]],
+    table: str,
+    column: str,
+) -> Optional[str]:
+    """Reads a vector column's declared DDL type, or None when it is not found."""
+    try:
+        rows = exec_fn("SELECT type FROM pragma_table_info(?) WHERE name = ?", [table, column])
+    except Exception:
+        return None
+    value = rows[0].get("type") if rows else None
+    return str(value) if value else None
+
+
+class SqliteVectorSupport:
+    """Registers the distance functions and loads sqlite-vec for a connection."""
+
+    def __init__(
+        self,
+        native: Optional[Callable[[], Any]] = None,
+        register_function: Optional[Callable[[str, Callable], None]] = None,
+        sqlite_vec: Optional[str] = None,
+    ):
+        self._native = native
+        self._register_function = register_function
+        self._sqlite_vec = sqlite_vec
+        self._capabilities: Optional[Dict[str, bool]] = None
+        # Declared column types and the "any binary row" probe, keyed by column.
+        # Both cost a query and neither changes while a schema is in use; a stale
+        # answer is still safe, because a strategy that fails falls through to the
+        # next one rather than reporting a wrong result.
+        self._column_types: Dict[str, Optional[str]] = {}
+        self._has_binary_rows: Dict[str, bool] = {}
+
+    def column_type(self, exec_fn, table: str, column: str) -> Optional[str]:
+        """The column's declared DDL type, remembered per column."""
+        key = f"{table}.{column}"
+        if key not in self._column_types:
+            self._column_types[key] = read_sqlite_column_type(exec_fn, table, column)
+        return self._column_types[key]
+
+    def has_binary_rows(self, exec_fn, table: str, column: str, tail: str, params: List) -> bool:
+        """Whether any row stores raw bytes, which `json_each` cannot read.
+
+        Checked only for the `sql` strategy, where a legacy JSON column that has
+        since taken new float32 rows would otherwise rank just the old ones.
+        """
+        clause = (tail + " AND ") if tail else " WHERE "
+        try:
+            rows = exec_fn(
+                f"SELECT 1 FROM {table}{clause}typeof({column}) = 'blob' LIMIT 1",
+                list(params))
+            return bool(rows)
+        except Exception:
+            return False
+
+    def capabilities(self) -> Dict[str, bool]:
+        """Probes once and caches the answer."""
+        if self._capabilities is not None:
+            return self._capabilities
+
+        caps = {"vec": False, "udf": False, "json1": False}
+        native = self._native() if self._native else None
+        if native is not None:
+            if self._sqlite_vec and hasattr(native, "enable_load_extension"):
+                try:
+                    native.enable_load_extension(True)
+                    native.load_extension(self._sqlite_vec)
+                except Exception:
+                    pass
+                finally:
+                    try:
+                        native.enable_load_extension(False)
+                    except Exception:
+                        pass
+            try:
+                native.execute("SELECT vec_version()").fetchone()
+                caps["vec"] = True
+            except Exception:
+                caps["vec"] = False
+            try:
+                native.execute("SELECT an5_vector_version()").fetchone()
+                caps["udf"] = True
+                caps["vec"] = False
+            except Exception:
+                pass
+
+        if self._register_function and not caps["udf"]:
+            try:
+                for metric, name in AN5_VECTOR_FUNCTIONS.items():
+                    self._register_function(name, make_distance_function(metric))
+                caps["udf"] = True
+            except Exception:
+                caps["udf"] = False
+
+        self._capabilities = caps
+        return caps
+
+    def probe(self, exec_fn: Callable[[str, List], List[Dict]]) -> Dict[str, bool]:
+        """Confirms JSON1 before a query is built that needs it."""
+        caps = self.capabilities()
+        if not caps["json1"]:
+            try:
+                exec_fn("SELECT json_valid('[1]') AS an5_json_ok", [])
+                caps["json1"] = True
+            except Exception:
+                caps["json1"] = False
+        return caps
+
+
+def run_sqlite_vector_search(
+    exec_fn: Callable[[str, List], List[Dict]],
+    table: str,
+    raw_table: str,
+    column: str,
+    raw_column: str,
+    vector: Sequence[float],
+    metric: str,
+    take: int,
+    tail: str = "",
+    params: Optional[List] = None,
+    support: Optional[SqliteVectorSupport] = None,
+    preference: Optional[str] = None,
+) -> Optional[List[Dict]]:
+    """Ranks a SQLite table inside the database.
+
+    Returns None when the connection offers no strategy, so the caller can fall
+    back to scoring in memory.
+    """
+    if vector is None or len(vector) == 0:
+        return None
+
+    if support is not None:
+        caps = support.probe(exec_fn)
+    else:
+        try:
+            exec_fn("SELECT json_valid('[1]') AS an5_json_ok", [])
+            caps = {"vec": False, "udf": False, "json1": True}
+        except Exception:
+            caps = {"vec": False, "udf": False, "json1": False}
+
+    declared = (support.column_type(exec_fn, raw_table, raw_column)
+                if support is not None
+                else read_sqlite_column_type(exec_fn, raw_table, raw_column))
+    for strategy in plan_sqlite_vector_strategies(caps, declared, preference):
+        if strategy == STRATEGY_SQL:
+            binary = (support.has_binary_rows(exec_fn, table, column, tail, list(params or []))
+                      if support is not None
+                      else bool(exec_fn("SELECT 1 FROM " + table + (tail + " AND " if tail else " WHERE ")
+                                        + f"typeof({column}) = 'blob' LIMIT 1", list(params or []))))
+            if binary:
+                continue
+        query, bound = build_sqlite_vector_query(
+            strategy, metric, table, column, vector, take, tail, list(params or [])
+        )
+        try:
+            return exec_fn(query, bound)
+        except Exception:
+            # The driver advertised the capability but the query failed, e.g. an
+            # extension that did not really load. The next strategy is cheaper
+            # than reporting the failure, and memory is the last resort.
+            continue
+    return None

package/python/sqlite/__init__.py

@@ -1,3 +1,3 @@
-from .provider import connect, is_memory, parse_connection_string, placeholder
+from .provider import connect, is_memory, parse_connection_string, placeholder, register_vector_functions
 
-__all__ = ["connect", "is_memory", "parse_connection_string", "placeholder"]
+__all__ = ["connect", "is_memory", "parse_connection_string", "placeholder", "register_vector_functions"]

package/python/sqlite/provider.py

@@ -12,7 +12,12 @@ statement) would leave transactions dangling. `None` means autocommit, matching
 """
 
 import sqlite3
-from typing import Tuple
+from typing import Optional, Tuple
+
+try:  # package import
+    from ..base.vectors import SqliteVectorSupport
+except ImportError:  # flat layout, when the folder itself is on sys.path
+    from base.vectors import SqliteVectorSupport
 
 # Strip the `sqlite:`/`sqlite://`/`sqlite:///` wrapper first, only then look for
 # `file:`. Two steps are required because `sqlite:file::memory:?cache=shared` only
@@ -79,5 +84,20 @@ def connect(connection_string: str):
     return conn
 
 
+def register_vector_functions(conn, sqlite_vec: Optional[str] = None) -> SqliteVectorSupport:
+    """Prepares the connection for in-database vector search.
+
+    Registers `an5_vec_cosine` / `an5_vec_l2` / `an5_vec_ip` so SQLite can rank
+    a `VECTOR(n)` column without the client loading it, and loads the sqlite-vec
+    extension when a path was given. A driver that refuses either is not an
+    error: the search then falls back to `json_each` or to Python.
+    """
+    return SqliteVectorSupport(
+        native=lambda: conn,
+        register_function=lambda name, fn: conn.create_function(name, 2, fn),
+        sqlite_vec=sqlite_vec,
+    )
+
+
 def placeholder() -> str:
     return "?"