# ApexBase API Reference
Complete API reference for ApexBase Python SDK.
## Table of Contents
1. [ApexClient](#apexclient) - Main client class
2. [ResultView](#resultview) - Query results
3. [Constants](#constants) - Module constants
4. [File Reading Table Functions](#file-reading-table-functions) - read_csv / read_parquet / read_json
5. [Set Operations](#set-operations) - UNION / INTERSECT / EXCEPT
6. [Vector Search](#vector-search) - topk_distance / batch_topk_distance / SQL explode_rename
---
## ApexClient
The main entry point for ApexBase operations.
### Constructor
```python
ApexClient(
dirpath: str = None,
batch_size: int = 1000,
drop_if_exists: bool = False,
enable_cache: bool = True,
cache_size: int = 10000,
prefer_arrow_format: bool = True,
durability: Literal['fast', 'safe', 'max'] = 'fast',
_auto_manage: bool = True
)
```
**Parameters:**
- `dirpath`: Data directory path (default: current directory)
- `batch_size`: Batch size for bulk operations
- `drop_if_exists`: If True, delete existing data on open
- `enable_cache`: Enable query result caching
- `cache_size`: Maximum cache entries
- `prefer_arrow_format`: Prefer Arrow format for internal transfers
- `durability`: Persistence level - 'fast' (async), 'safe' (sync), 'max' (fsync every write)
**Example:**
```python
from apexbase import ApexClient
# Basic usage
client = ApexClient("./data")
# With durability options
client = ApexClient("./data", durability="safe")
# Clean start (drop existing)
client = ApexClient.create_clean("./data")
```
---
### Database Management
ApexBase supports multiple isolated databases within a single root directory. Each named database is stored as a subdirectory; the `'default'` database maps to the root directory (backward-compatible).
#### use_database
```python
use_database(database: str = 'default') -> ApexClient
```
Switch to a named database. Creates the database subdirectory if it does not exist. Resets the current table to `None`.
**Parameters:**
- `database`: Database name. `'default'` (or `''`) maps to the root directory.
**Returns:** `self` (for method chaining)
**Examples:**
```python
# Switch to analytics database
client.use_database("analytics")
# Switch back to default (root-level tables)
client.use_database("default")
# Method chaining
client.use_database("hr").create_table("employees")
```
---
#### use
```python
use(database: str = 'default', table: str = None) -> ApexClient
```
Switch to a named database and optionally select or create a table in one call. If `table` is specified and does not exist it is created automatically.
**Parameters:**
- `database`: Database name (default = root-level).
- `table`: Table name to select. If `None`, only the database is switched.
**Returns:** `self` (for method chaining)
**Examples:**
```python
# Switch database only
client.use(database="analytics")
# Switch database and select an existing table
client.use(database="analytics", table="events")
# Switch database and auto-create table if missing
client.use(database="new_db", table="new_table")
client.store({"key": "value"})
```
---
#### list_databases
```python
list_databases() -> List[str]
```
Return a sorted list of all available databases. `'default'` is always included.
**Example:**
```python
dbs = client.list_databases()
print(dbs) # ['analytics', 'default', 'hr']
```
---
#### current_database
```python
current_database: str # Property
```
Return the name of the currently active database. Returns `'default'` when operating on root-level tables.
**Example:**
```python
client.use_database("analytics")
print(client.current_database) # 'analytics'
```
---
### Cross-Database SQL
All SQL operations support the standard `database.table` qualified name syntax. The active database context only affects unqualified table references; qualified references always resolve to the correct database regardless of context.
**Supported operations:**
```python
# SELECT across databases
client.execute("SELECT * FROM default.users")
client.execute("SELECT * FROM analytics.events WHERE cnt > 10")
# JOIN across databases
client.execute("""
SELECT u.name, e.event
FROM default.users u
JOIN analytics.events e ON u.id = e.user_id
""")
# INSERT into a different database
client.execute("INSERT INTO analytics.events (name, cnt) VALUES ('click', 1)")
# UPDATE in a different database
client.execute("UPDATE default.users SET age = 31 WHERE name = 'Alice'")
# DELETE from a different database
client.execute("DELETE FROM default.users WHERE age < 18")
# DDL across databases
client.execute("CREATE TABLE analytics.summary (total INT)")
client.execute("DROP TABLE IF EXISTS analytics.old_table")
```
---
### Table Management
#### create_table
```python
create_table(table_name: str, schema: dict = None) -> None
```
Create a new table, optionally with a pre-defined schema.
**Parameters:**
- `table_name`: Name of the table to create.
- `schema`: Optional dict mapping column names to type strings. Pre-defining schema avoids type inference on the first insert, providing a performance benefit for bulk loading.
**Supported types:** `int8`, `int16`, `int32`, `int64`, `uint8`, `uint16`, `uint32`, `uint64`, `float32`, `float64`, `bool`, `string`, `binary`
**Examples:**
```python
# Without schema
client.create_table("users")
# With pre-defined schema
client.create_table("orders", schema={
"order_id": "int64",
"product": "string",
"price": "float64",
"paid": "bool"
})
```
#### drop_table
```python
drop_table(table_name: str) -> None
```
Drop a table and all its data.
**Example:**
```python
client.drop_table("old_table")
```
#### use_table
```python
use_table(table_name: str) -> None
```
Switch to a different table for subsequent operations.
**Example:**
```python
client.use_table("users")
```
#### list_tables
```python
list_tables() -> List[str]
```
Return list of all table names.
**Example:**
```python
tables = client.list_tables()
print(tables) # ['users', 'orders']
```
#### current_table
```python
current_table: Optional[str] # Property
```
Get the name of the currently active table. Returns `None` if no table is selected.
**Example:**
```python
print(client.current_table) # 'users'
```
---
### Data Storage
#### store
```python
store(data) -> None
```
Store data in the active table. Requires a table to be selected via `create_table()` or `use_table()` first. Accepts multiple formats:
- Single dict: `{"name": "Alice", "age": 30}`
- List of dicts: `[{"name": "A"}, {"name": "B"}]`
- Dict of columns: `{"name": ["A", "B"], "age": [20, 30]}`
- pandas DataFrame
- polars DataFrame
- PyArrow Table
**Examples:**
```python
# Single record
client.store({"name": "Alice", "age": 30})
# Multiple records
client.store([
{"name": "Bob", "age": 25},
{"name": "Charlie", "age": 35}
])
# Columnar format (fastest for bulk)
client.store({
"name": ["David", "Eve"],
"age": [28, 32]
})
```
#### from_pandas
```python
from_pandas(df: pd.DataFrame, table_name: str = None) -> ApexClient
```
Import data from pandas DataFrame. Returns self for chaining.
**Parameters:**
- `df`: pandas DataFrame to import
- `table_name`: Optional. If provided, auto-creates/selects the table before importing.
**Example:**
```python
import pandas as pd
df = pd.DataFrame({"name": ["A", "B"], "age": [20, 30]})
client.from_pandas(df, table_name="users")
```
#### from_polars
```python
from_polars(df: pl.DataFrame, table_name: str = None) -> ApexClient
```
Import data from polars DataFrame. Returns self for chaining.
**Parameters:**
- `df`: polars DataFrame to import
- `table_name`: Optional. If provided, auto-creates/selects the table before importing.
**Example:**
```python
import polars as pl
df = pl.DataFrame({"name": ["A", "B"], "age": [20, 30]})
client.from_polars(df, table_name="users")
```
#### from_pyarrow
```python
from_pyarrow(table: pa.Table, table_name: str = None) -> ApexClient
```
Import data from PyArrow Table. Returns self for chaining.
**Parameters:**
- `table`: PyArrow Table to import
- `table_name`: Optional. If provided, auto-creates/selects the table before importing.
**Example:**
```python
import pyarrow as pa
table = pa.table({"name": ["A", "B"], "age": [20, 30]})
client.from_pyarrow(table, table_name="users")
```
---
### Data Retrieval
#### execute
```python
execute(sql: str, show_internal_id: bool = None) -> ResultView
```
Execute SQL query and return results.
**Parameters:**
- `sql`: SQL statement (SELECT, INSERT, etc.)
- `show_internal_id`: If True, include _id column in results
**Example:**
```python
results = client.execute("SELECT * FROM users WHERE age > 25")
results = client.execute("SELECT COUNT(*), AVG(age) FROM users")
count = results.scalar()
results = client.execute("SELECT name, age FROM users ORDER BY age DESC LIMIT 10")
```
#### query
```python
query(
sql: str = None,
where_clause: str = None,
limit: int = None
) -> ResultView
```
Query with WHERE expression (backward compatibility).
**Example:**
```python
# WHERE expression only
results = client.query("age > 25")
results = client.query("name LIKE 'A%'")
# With limit
results = client.query(where_clause="city = 'NYC'", limit=100)
```
#### retrieve
```python
retrieve(id_: int) -> Optional[dict]
```
Get a single record by its internal _id.
**Example:**
```python
record = client.retrieve(0)
print(record) # {'_id': 0, 'name': 'Alice', 'age': 30}
```
#### retrieve_many
```python
retrieve_many(ids: List[int]) -> ResultView
```
Get multiple records by their internal _ids.
**Example:**
```python
results = client.retrieve_many([0, 1, 2, 5])
df = results.to_pandas()
```
#### retrieve_all
```python
retrieve_all() -> ResultView
```
Get all records from the current table.
**Example:**
```python
results = client.retrieve_all()
print(len(results)) # Total row count
```
#### count_rows
```python
count_rows(table_name: str = None) -> int
```
Count rows in a table.
**Example:**
```python
count = client.count_rows()
count = client.count_rows("users") # Specific table
```
---
### Data Modification
#### replace
```python
replace(id_: int, data: dict) -> bool
```
Replace a record by _id.
**Example:**
```python
success = client.replace(0, {"name": "Alice", "age": 31})
```
#### batch_replace
```python
batch_replace(data_dict: Dict[int, dict]) -> List[int]
```
Batch replace multiple records.
**Example:**
```python
updated = client.batch_replace({
0: {"name": "Alice", "age": 31},
1: {"name": "Bob", "age": 26}
})
```
#### delete
```python
delete(ids: Union[int, List[int]]) -> bool
```
Delete record(s) by _id.
**Example:**
```python
# Single delete
client.delete(5)
# Batch delete
client.delete([1, 2, 3])
```
---
### Column Operations
#### add_column
```python
add_column(column_name: str, column_type: str) -> None
```
Add a new column to the current table.
**Types:** `Int8`, `Int16`, `Int32`, `Int64`, `UInt8`, `UInt16`, `UInt32`, `UInt64`, `Float32`, `Float64`, `String`, `Bool`
**Example:**
```python
client.add_column("email", "String")
client.add_column("score", "Float64")
```
#### drop_column
```python
drop_column(column_name: str) -> None
```
Drop a column from the current table. Cannot drop _id column.
**Example:**
```python
client.drop_column("temp_field")
```
#### rename_column
```python
rename_column(old_column_name: str, new_column_name: str) -> None
```
Rename a column. Cannot rename _id column.
**Example:**
```python
client.rename_column("email", "email_address")
```
#### get_column_dtype
```python
get_column_dtype(column_name: str) -> str
```
Get the data type of a column.
**Example:**
```python
dtype = client.get_column_dtype("age") # 'Int64'
```
#### list_fields
```python
list_fields() -> List[str]
```
List all column names in the current table.
**Example:**
```python
fields = client.list_fields()
print(fields) # ['_id', 'name', 'age', 'city']
```
---
### Full-Text Search
FTS is implemented natively in Rust and available through all interfaces (Python API, PG Wire, Arrow Flight). The recommended way to manage and query FTS indexes is via SQL.
#### FTS SQL DDL and Query Reference
| `CREATE FTS INDEX ON table (col1, col2)` | Create FTS index on specified columns |
| `CREATE FTS INDEX ON table` | Create FTS index on all string columns |
| `CREATE FTS INDEX ON table WITH (opt=val)` | Create with options |
| `DROP FTS INDEX ON table` | Drop index and delete files |
| `ALTER FTS INDEX ON table DISABLE` | Suspend indexing, keep files |
| `ALTER FTS INDEX ON table ENABLE` | Resume indexing and back-fill any missed rows |
| `SHOW FTS INDEXES` | List FTS-configured tables across all databases |
| `WHERE MATCH('query')` | Exact / ranked full-text search |
| `WHERE FUZZY_MATCH('query')` | Fuzzy / typo-tolerant search |
**`CREATE FTS INDEX`**
```sql
CREATE FTS INDEX ON table_name [(col1, col2, ...)] [WITH (lazy_load=bool, cache_size=N)]
```
- `(col1, col2)` — optional column list; omit to index all string columns
- `lazy_load` — defer loading index into RAM until first search (default `false`)
- `cache_size` — LRU cache entries for the index (default `10000`)
```python
client.execute("CREATE FTS INDEX ON articles (title, content)")
client.execute("CREATE FTS INDEX ON logs WITH (lazy_load=true, cache_size=50000)")
```
**`DROP FTS INDEX`**
```sql
DROP FTS INDEX ON table_name
```
Removes the index entry and deletes the `.nfts` index files from disk.
**`ALTER FTS INDEX ... DISABLE`**
```sql
ALTER FTS INDEX ON table_name DISABLE
```
Suspends FTS write-sync (INSERT / DELETE no longer update the index). Index files are kept on disk. Use `ALTER FTS INDEX ... ENABLE` to resume.
**`ALTER FTS INDEX ... ENABLE`**
```sql
ALTER FTS INDEX ON table_name ENABLE
```
Resumes FTS write-sync and **back-fills all rows** currently in the table, including any rows inserted while FTS was disabled.
**`SHOW FTS INDEXES`**
```sql
SHOW FTS INDEXES
```
Returns a result set with columns: `database`, `table`, `enabled`, `fields`, `lazy_load`, `cache_size`. Lists indexes across the root directory and all named sub-databases.
```python
result = client.execute("SHOW FTS INDEXES")
df = result.to_pandas()
# database table enabled fields lazy_load cache_size
# default articles True title, content False 10000
```
**`MATCH('query')`**
Used in `WHERE` clauses to filter rows whose indexed text contains all query terms.
```python
# Simple search
client.execute("SELECT * FROM articles WHERE MATCH('python tutorial')")
# Combined with other conditions
client.execute("""
SELECT title, content FROM articles
WHERE MATCH('machine learning') AND year >= 2023
ORDER BY _id DESC
LIMIT 10
""")
# Aggregations
client.execute("SELECT COUNT(*) FROM articles WHERE MATCH('rust')")
```
**`FUZZY_MATCH('query')`**
Like `MATCH()` but tolerates typos and spelling variations.
```python
client.execute("SELECT * FROM articles WHERE FUZZY_MATCH('pytohn')") # matches 'python'
client.execute("SELECT * FROM articles WHERE FUZZY_MATCH('databse')") # matches 'database'
```
> **Note:** `MATCH()` / `FUZZY_MATCH()` require a FTS index to exist for the queried table. Use `CREATE FTS INDEX ON table` first. The SQL interface works over all transports (Python, PG Wire, Arrow Flight).
---
#### Python API
#### init_fts
```python
init_fts(
table_name: str = None,
index_fields: Optional[List[str]] = None,
lazy_load: bool = False,
cache_size: int = 10000
) -> ApexClient
```
Initialize full-text search for a table.
**Parameters:**
- `table_name`: Table to index (default: current table)
- `index_fields`: Fields to index (None = all string fields)
- `lazy_load`: Load index on first search
- `cache_size`: FTS cache size
**Example:**
```python
client.init_fts(index_fields=["title", "content"])
client.init_fts(index_fields=["name"], lazy_load=True)
```
#### search_text
```python
search_text(query: str, table_name: str = None) -> np.ndarray
```
Search for documents containing query terms. Returns array of _ids.
**Example:**
```python
ids = client.search_text("database")
print(ids) # array([0, 5, 10])
```
#### fuzzy_search_text
```python
fuzzy_search_text(
query: str,
min_results: int = 1,
table_name: str = None
) -> np.ndarray
```
Fuzzy search tolerating typos. Returns array of _ids.
**Example:**
```python
ids = client.fuzzy_search_text("databse") # Matches "database"
```
#### search_and_retrieve
```python
search_and_retrieve(
query: str,
table_name: str = None,
limit: Optional[int] = None,
offset: int = 0
) -> ResultView
```
Search and return full records.
**Example:**
```python
results = client.search_and_retrieve("python")
results = client.search_and_retrieve("python", limit=10, offset=20)
```
#### search_and_retrieve_top
```python
search_and_retrieve_top(
query: str,
n: int = 100,
table_name: str = None
) -> ResultView
```
Return top N search results.
**Example:**
```python
results = client.search_and_retrieve_top("important", n=5)
```
#### get_fts_stats
```python
get_fts_stats(table_name: str = None) -> Dict
```
Get FTS statistics.
**Example:**
```python
stats = client.get_fts_stats()
print(stats) # {'fts_enabled': True, 'doc_count': 1000, 'term_count': 5000}
```
#### disable_fts
```python
disable_fts(table_name: str = None) -> ApexClient
```
Disable FTS (keeps index files).
**Example:**
```python
client.disable_fts()
```
#### drop_fts
```python
drop_fts(table_name: str = None) -> ApexClient
```
Disable FTS and delete index files.
**Example:**
```python
client.drop_fts()
```
---
### Utility Methods
#### flush
```python
flush() -> None
```
Flush all pending writes to disk.
**Example:**
```python
client.flush()
```
#### set_auto_flush
```python
set_auto_flush(rows: int = 0, bytes: int = 0) -> None
```
Set auto-flush thresholds.
**Example:**
```python
client.set_auto_flush(rows=1000) # Flush every 1000 rows
client.set_auto_flush(bytes=1024*1024) # Flush every 1MB
```
#### get_auto_flush
```python
get_auto_flush() -> tuple
```
Get current auto-flush configuration.
**Example:**
```python
rows, bytes = client.get_auto_flush()
```
#### estimate_memory_bytes
```python
estimate_memory_bytes() -> int
```
Estimate current memory usage in bytes.
**Example:**
```python
mem_bytes = client.estimate_memory_bytes()
print(f"Using {mem_bytes / 1024 / 1024:.1f} MB")
```
#### close
```python
close() -> None
```
Close the client and release resources.
**Example:**
```python
client.close()
```
---
## ResultView
Container for query results with multiple output formats.
### Conversion Methods
#### to_pandas
```python
to_pandas(zero_copy: bool = True) -> pd.DataFrame
```
Convert to pandas DataFrame.
**Parameters:**
- `zero_copy`: Use ArrowDtype for zero-copy (pandas 2.0+)
**Example:**
```python
df = results.to_pandas()
df = results.to_pandas(zero_copy=False) # Traditional NumPy types
```
#### to_polars
```python
to_polars() -> pl.DataFrame
```
Convert to polars DataFrame.
**Example:**
```python
df = results.to_polars()
```
#### to_arrow
```python
to_arrow() -> pa.Table
```
Convert to PyArrow Table.
**Example:**
```python
table = results.to_arrow()
```
#### to_dict
```python
to_dict() -> List[dict]
```
Convert to list of dictionaries.
**Example:**
```python
records = results.to_dict()
for record in records:
print(record["name"])
```
### Access Methods
#### scalar
```python
scalar() -> Any
```
Get single scalar value (for aggregate queries).
**Example:**
```python
count = client.execute("SELECT COUNT(*) FROM users").scalar()
```
#### first
```python
first() -> Optional[dict]
```
Get first row as dictionary.
**Example:**
```python
row = results.first()
```
#### get_ids
```python
get_ids(return_list: bool = False) -> Union[np.ndarray, List[int]]
```
Get internal _ids from results.
**Example:**
```python
ids = results.get_ids() # numpy array (default)
ids = results.get_ids(return_list=True) # Python list
```
### Properties
#### shape
```python
shape: tuple # (rows, columns)
```
#### columns
```python
columns: List[str]
```
### Sequence Interface
```python
# Length
len(results)
# Iteration
for row in results:
print(row)
# Indexing
first = results[0]
second = results[1]
```
---
## Constants
### Module Constants
```python
from apexbase import (
__version__, # Package version
FTS_AVAILABLE, # True (FTS always available)
ARROW_AVAILABLE, # True if pyarrow installed
POLARS_AVAILABLE, # True if polars installed
DurabilityLevel, # Type hint: Literal['fast', 'safe', 'max']
)
```
---
## SQL Support
ApexBase supports standard SQL for querying.
### Quoted Identifiers
When a column name collides with a SQL reserved keyword (e.g. `order`, `group`, `select`, `table`), wrap it in **backticks** (Hive/MySQL style) or **double quotes** (SQL standard) so the parser treats it as an identifier instead of a keyword.
| Backtick (Hive/MySQL) | `` `column` `` | `` SELECT `order`, `group` FROM t `` |
| Double-quote (SQL standard) | `"column"` | `SELECT "order", "group" FROM t` |
Both styles produce identical results — use whichever you prefer.
**Examples:**
```sql
-- Backtick style
SELECT `order`, `group`, `select` FROM orders WHERE `order` > 100 ORDER BY `order` DESC
-- Double-quote style
SELECT "order", "group", "select" FROM orders WHERE "order" > 100 ORDER BY "order" DESC
-- Mixed (both styles in one query)
SELECT `order`, "group" FROM orders
-- Works in all clauses: SELECT, WHERE, ORDER BY, GROUP BY, HAVING, INSERT, etc.
INSERT INTO t (`order`, `group`) VALUES (1, 'A')
SELECT `group`, COUNT(*) FROM t GROUP BY `group` HAVING COUNT(*) > 5
```
> **Tip:** Quoting is only needed when the column name is a reserved keyword. Regular column names like `name`, `age`, `city` do not need quoting.
### SELECT
```sql
SELECT * FROM table
SELECT col1, col2 FROM table
SELECT col1 AS alias FROM table
SELECT DISTINCT col1 FROM table
SELECT * FROM table WHERE condition
SELECT * FROM table ORDER BY col DESC
SELECT * FROM table LIMIT 100
SELECT * FROM table LIMIT 100 OFFSET 10
SELECT * FROM table ORDER BY col LIMIT 100
```
### Aggregate Functions
```sql
SELECT COUNT(*) FROM table
SELECT COUNT(DISTINCT col) FROM table
SELECT SUM(col), AVG(col), MAX(col), MIN(col) FROM table
```
### WHERE Clauses
```sql
WHERE col = value
WHERE col > value
WHERE col LIKE 'pattern%'
WHERE col IN (1, 2, 3)
WHERE col BETWEEN 10 AND 20
WHERE col IS NULL
WHERE col IS NOT NULL
WHERE condition1 AND condition2
WHERE condition1 OR condition2
```
### GROUP BY / HAVING
```sql
SELECT category, COUNT(*), AVG(price)
FROM products
GROUP BY category
SELECT category, COUNT(*)
FROM products
GROUP BY category
HAVING COUNT(*) > 10
```
### JOINs
```sql
SELECT * FROM table1 JOIN table2 ON table1.id = table2.id
```
---
## File Reading Table Functions
ApexBase provides SQL table functions that read external files directly in a `FROM` clause without importing data into a table first. All three functions return a result compatible with the full SQL engine — you can apply `WHERE`, `GROUP BY`, `ORDER BY`, `LIMIT`, `JOIN`, and `UNION` on top of them.
### read_csv
```sql
SELECT * FROM read_csv('path/to/file.csv')
SELECT * FROM read_csv('path/to/file.csv', header=true, delimiter=',')
```
**Parameters:**
| `header` | `true` | Whether the first row is a header. Set to `false` / `0` if the file has no header row. |
| `delimiter` / `delim` / `sep` | `,` | Field delimiter character. Use `'\t'` for TSV files. |
**Schema inference**: types are inferred automatically from the first 100 data rows (Int64, Float64, Bool, or String).
**Examples:**
```python
# Read a comma-delimited CSV with header row
result = client.execute("SELECT * FROM read_csv('/data/sales.csv')")
# Tab-separated values (TSV)
result = client.execute("SELECT * FROM read_csv('/data/data.tsv', delimiter='\t')")
# No header row
result = client.execute("SELECT * FROM read_csv('/data/raw.csv', header=false)")
# Full SQL on top of the file
result = client.execute("""
SELECT city, COUNT(*) AS cnt, AVG(price)
FROM read_csv('/data/orders.csv')
WHERE price > 100
GROUP BY city
ORDER BY cnt DESC
LIMIT 10
""")
# Convert directly to DataFrame
df = client.execute("SELECT * FROM read_csv('/data/large.csv')").to_pandas()
```
### read_parquet
```sql
SELECT * FROM read_parquet('path/to/file.parquet')
```
No options — schema is read directly from the Parquet file's metadata.
**Examples:**
```python
# Read a Parquet file
result = client.execute("SELECT * FROM read_parquet('/data/events.parquet')")
# Projection and filter
result = client.execute("""
SELECT user_id, SUM(amount) AS total
FROM read_parquet('/data/transactions.parquet')
WHERE category = 'food'
GROUP BY user_id
""")
# Zero-copy to Arrow
table = client.execute("SELECT * FROM read_parquet('/data/wide.parquet')").to_arrow()
```
### read_json
```sql
SELECT * FROM read_json('path/to/file.json')
```
Handles two formats automatically:
- **NDJSON / JSON Lines** — one JSON object per line (`.json`, `.jsonl`, `.ndjson`)
- **pandas column-oriented JSON** — output of `df.to_json(orient='columns')` or `orient='split'`
No options — format is detected automatically.
**Examples:**
```python
# Read NDJSON (one JSON object per line)
result = client.execute("SELECT * FROM read_json('/data/logs.ndjson')")
# Read pandas-exported JSON
result = client.execute("SELECT * FROM read_json('/data/export.json')")
# Apply aggregation
result = client.execute("""
SELECT status, COUNT(*) AS cnt
FROM read_json('/data/events.json')
GROUP BY status
ORDER BY cnt DESC
""")
```
### Joining file reads with tables
```python
# JOIN a file read with a stored table
result = client.execute("""
SELECT u.name, f.score
FROM users u
JOIN read_csv('/data/scores.csv') f ON u.id = f.user_id
WHERE f.score > 90
""")
# UNION a file with a table
result = client.execute("""
SELECT name, city FROM users
UNION ALL
SELECT name, city FROM read_csv('/data/new_users.csv')
""")
```
### Performance notes
- All three functions use `mmap` for zero-copy file access.
- CSV and JSON files are parsed in parallel (one chunk per CPU core via Rayon).
- Parquet files use parallel column decoding with shared metadata (zero re-parse overhead).
- Benchmarked against Polars on 1M rows: CSV 0.95×, NDJSON 0.93×, Parquet 1.33× (Arrow output).
---
## Set Operations
SQL set operations combine result sets from two or more `SELECT` statements. Both sides must return the same number of columns.
### UNION
Combines rows from both sides and **removes duplicate rows**.
```sql
SELECT col FROM table_a
UNION
SELECT col FROM table_b
```
### UNION ALL
Combines rows from both sides and **keeps all rows**, including duplicates.
```sql
SELECT col FROM table_a
UNION ALL
SELECT col FROM table_b
```
### INTERSECT
Returns only rows that appear in **both** result sets (deduplicated).
```sql
SELECT col FROM table_a
INTERSECT
SELECT col FROM table_b
```
### EXCEPT
Returns rows from the **left** side that do **not** appear in the right side (deduplicated).
```sql
SELECT col FROM table_a
EXCEPT
SELECT col FROM table_b
```
### Ordering and limiting set operation results
Append `ORDER BY`, `LIMIT`, and `OFFSET` after the final set operation; they apply to the combined result:
```sql
SELECT val FROM a
UNION
SELECT val FROM b
ORDER BY val ASC
LIMIT 100
```
### Examples
```python
# UNION: unique names across two tables
client.execute("""
SELECT name FROM customers
UNION
SELECT name FROM leads
ORDER BY name
""")
# UNION ALL: all rows from both tables (duplicates kept)
client.execute("""
SELECT name, city FROM domestic_users
UNION ALL
SELECT name, city FROM international_users
""")
# INTERSECT: users who appear in both the orders and wishlist tables
client.execute("""
SELECT user_id FROM orders
INTERSECT
SELECT user_id FROM wishlist
""")
# EXCEPT: users who placed orders but have no open support tickets
client.execute("""
SELECT user_id FROM orders
EXCEPT
SELECT user_id FROM support_tickets WHERE status = 'open'
""")
# Works with read_csv too
client.execute("""
SELECT email FROM users
EXCEPT
SELECT email FROM read_csv('/data/unsubscribed.csv')
""")
```
### Set operation summary
| `UNION` | removed | left ∪ right (unique) |
| `UNION ALL` | kept | all rows from both sides |
| `INTERSECT` | removed | left ∩ right |
| `EXCEPT` | removed | left \ right |
---
## Vector Search
ApexBase has a built-in vector similarity search engine implemented entirely in Rust with SIMD-accelerated distance kernels and an OS-level mmap scan buffer that is populated once and reused across queries. Both single-query and batch modes are available through dedicated Python methods and through a SQL extension syntax.
### Vector column storage
Vectors are stored as **FixedList** columns when inserted as numpy arrays (recommended) or as **Binary** columns when inserted as Python lists or tuples.
```python
import numpy as np
from apexbase import ApexClient
client = ApexClient("./vecdb")
client.create_table("items")
# numpy arrays → FixedList column (optimal, zero-copy mmap scan)
client.store({
"label": ["a", "b", "c"],
"vec": [np.random.rand(128).astype(np.float32) for _ in range(3)],
})
# Python list/tuple → Binary column (also supported)
client.store({"label": "d", "vec": [0.1, 0.2, 0.3]})
```
---
### Float16 Vector Storage (`FLOAT16_VECTOR`)
ApexBase supports **half-precision (float16) vector columns** for memory-efficient embedding storage. Each element is stored as a 16-bit IEEE 754 half-precision float, halving the memory and I/O footprint compared to float32 vectors.
#### Declaring a float16 column
```sql
CREATE TABLE embeddings (id TEXT, vec FLOAT16_VECTOR)
```
Accepted type name aliases: `FLOAT16_VECTOR`, `FLOAT16VECTOR`, `F16_VECTOR`.
#### Inserting float16 vectors
Use a **batch store** (two or more records in one call) for optimal encoding. Pass vectors as `numpy` arrays — any numeric dtype is accepted; the storage layer converts to float16 automatically.
```python
import numpy as np
from apexbase import ApexClient
client = ApexClient("./vecdb")
client.execute("CREATE TABLE embeddings (label TEXT, vec FLOAT16_VECTOR)")
client.use_table("embeddings")
# float32 source data — auto-quantized to f16 on write
vecs = np.random.rand(1000, 128).astype(np.float32)
client.store([{"label": str(i), "vec": vecs[i]} for i in range(len(vecs))])
# float16 source data — stored directly
vecs_f16 = vecs.astype(np.float16)
client.store([{"label": str(i), "vec": vecs_f16[i]} for i in range(len(vecs_f16))])
```
> **Note:** Always use batch store (`len(data) > 1`) for `FLOAT16_VECTOR` columns. The single-record path converts vectors to float32 bytes before storage, which will produce incorrect distances on float16 columns. Batch the writes or use the columnar dict API:
>
> ```python
> # Columnar dict — also correct and fastest
> client.store({"label": [str(i) for i in range(n)], "vec": [vecs[i] for i in range(n)]})
> ```
#### Querying float16 vectors
All `topk_distance`, `batch_topk_distance`, and SQL distance functions work transparently on `FLOAT16_VECTOR` columns. The query vector is always provided in float32/float64 — no special handling needed.
```python
query = np.random.rand(128).astype(np.float32)
# TopK search
results = client.topk_distance("vec", query, k=10, metric="l2")
# SQL distance functions also work
client.execute("""
SELECT label, array_distance(vec, [0.1, 0.2, 0.3, 0.4]) AS dist
FROM embeddings
ORDER BY dist
LIMIT 5
""")
```
#### SIMD acceleration
Float16 distance kernels are hardware-accelerated on supported CPUs:
| `aarch64` (Apple M-series, AWS Graviton) | `fp16` NEON | L2, Dot, Cosine, L1, L∞ — FCVTL/FCVTL2 |
| `x86_64` | `f16c` + `AVX2` | L2, Dot, Cosine, L1, L∞ — `_cvtph_ps` |
| All others | scalar fallback | all metrics |
CPU feature detection is automatic at runtime — no build flags or environment variables required. On Apple Silicon (M1/M2/M3/M4), f16 kernels are typically **≥2× faster** than equivalent float32 kernels.
#### Quantization error
Float16 has ~3 decimal digits of precision (machine epsilon ≈ 9.77 × 10⁻⁴). For unit vectors or typical embedding ranges [−1, 1], the relative distance error is under 0.2%.
```python
import numpy as np
def f16_quantize(v):
return v.astype(np.float16).astype(np.float32)
vec = np.random.rand(128).astype(np.float32)
q = np.random.rand(128).astype(np.float32)
exact = float(np.sqrt(np.sum((vec - q) ** 2)))
f16_dist = float(np.sqrt(np.sum((f16_quantize(vec) - q) ** 2)))
print(f"relative error: {abs(exact - f16_dist) / exact:.2e}") # typically < 2e-3
```
---
### topk_distance
```python
topk_distance(
col: str,
query,
k: int = 10,
metric: str = 'l2',
id_col: str = '_id',
dist_col: str = 'dist',
) -> ResultView
```
Heap-based nearest-neighbour search: O(n log k), significantly faster than `ORDER BY distance LIMIT k` for large tables.
**Parameters:**
- `col`: Name of the vector column to search (FixedList or Binary).
- `query`: Query vector — list, tuple, or numpy array of floats.
- `k`: Number of nearest neighbours to return (default `10`).
- `metric`: Distance metric (see table below).
- `id_col`: Column name for the returned row IDs (default `'_id'`).
- `dist_col`: Column name for the returned distances (default `'dist'`).
**Supported metrics:**
| `'l2'` | `'euclidean'` | √Σ(aᵢ−bᵢ)² |
| `'l2_squared'` | — | Σ(aᵢ−bᵢ)² |
| `'l1'` | `'manhattan'` | Σ|aᵢ−bᵢ| |
| `'linf'` | `'chebyshev'` | max|aᵢ−bᵢ| |
| `'cosine'` | `'cosine_distance'` | 1 − (a·b)/(‖a‖‖b‖) |
| `'dot'` | `'inner_product'` | −(a·b) (negated for min-heap) |
**Returns:** `ResultView` with `id_col` (Int64) and `dist_col` (Float64) columns, sorted nearest first.
**Example:**
```python
import numpy as np
query = np.random.rand(128).astype(np.float32)
# L2 (default)
results = client.topk_distance('vec', query, k=10)
df = results.to_pandas()
# df columns: _id (int64), dist (float64)
# Cosine distance, custom column names
results = client.topk_distance('vec', query, k=5, metric='cosine',
id_col='item_id', dist_col='cosine_dist')
# Join back to the original table to retrieve full records
top_ids = results.get_ids() # numpy array of _id values
records = client.retrieve_many(top_ids.tolist())
# Or use a SQL subquery
client.execute("""
SELECT items.label, items.vec
FROM items
WHERE _id IN (
SELECT _id FROM (
SELECT explode_rename(topk_distance(vec, [0.1, 0.2, 0.3], 5, 'l2'), '_id', 'dist')
FROM items
)
)
""")
```
---
### batch_topk_distance
```python
batch_topk_distance(
col: str,
queries,
k: int = 10,
metric: str = 'l2',
) -> numpy.ndarray
```
Batch nearest-neighbour search — N query vectors in a single Rust call.
**Why use this instead of calling `topk_distance` N times:**
- The mmap float buffer (`scan_buf`) is populated **once** regardless of N.
- All N queries run in **parallel** via Rayon (outer parallelism over queries).
- The `_id` column is read only once.
**Parameters:**
- `col`: Name of the vector column (FixedList or Binary).
- `queries`: `(N, D)` numpy array or array-like of query vectors (float32 or float64). A 1-D array is treated as a single query (N=1).
- `k`: Number of nearest neighbours per query (default `10`).
- `metric`: Distance metric — same values accepted as `topk_distance`.
**Returns:** `numpy.ndarray` of shape `(N, k, 2)`, dtype `float64`.
- `result[i, j, 0]` — `_id` of the j-th nearest neighbour for query i (cast to `int64` as needed).
- `result[i, j, 1]` — corresponding distance.
- Each row is sorted ascending by distance.
- Entries padded with `(-1, inf)` when fewer than k neighbours exist.
**Example:**
```python
import numpy as np
N, D = 100, 128
queries = np.random.rand(N, D).astype(np.float32)
result = client.batch_topk_distance('vec', queries, k=10)
# result.shape == (100, 10, 2)
ids = result[:, :, 0].astype(np.int64) # shape (100, 10)
dists = result[:, :, 1] # shape (100, 10)
# Nearest neighbour for each query
nearest_id = ids[:, 0] # shape (100,)
nearest_dist = dists[:, 0] # shape (100,)
# Cosine similarity batch search
result_cos = client.batch_topk_distance('vec', queries, k=5, metric='cosine')
```
---
### SQL: `explode_rename(topk_distance(...))`
Vector search is also available as a pure SQL expression. This is the form used internally by `topk_distance()` and is useful when composing larger SQL queries.
**Syntax:**
```sql
SELECT explode_rename(
topk_distance(col, [q1, q2, ..., qD], k, 'metric'),
'id_column_name',
'dist_column_name'
)
FROM table_name
```
- `col` — vector column name.
- `[q1, q2, ..., qD]` — query vector as an array literal (float values).
- `k` — integer number of results.
- `'metric'` — distance metric string.
- The two string arguments to `explode_rename` name the output columns.
`explode_rename` "explodes" the TopK pairs returned by `topk_distance` into k rows with two named columns.
**Examples:**
```python
# Basic: top 10 by L2 distance
results = client.execute("""
SELECT explode_rename(topk_distance(vec, [0.1, 0.2, 0.3], 10, 'l2'), '_id', 'dist')
FROM items
""")
df = results.to_pandas()
# df: _id (int64), dist (float64), 10 rows sorted nearest first
# Cosine distance with custom column names
results = client.execute("""
SELECT explode_rename(
topk_distance(vec, [1.0, 0.0, 0.0], 5, 'cosine'),
'item_id', 'cosine_dist'
)
FROM items
""")
# Dot product (inner product) search
results = client.execute("""
SELECT explode_rename(topk_distance(vec, [0.5, 0.5, 0.5], 20, 'dot'), '_id', 'score')
FROM embeddings
""")
```
> **Note:** The SQL `topk_distance` / `explode_rename` syntax requires the array literal `[...]` form for the query vector. To use dynamic vectors from Python, use the `topk_distance()` method instead, which handles the formatting automatically.
---
### Vector Search Performance
Benchmark: 1M rows × dim=128, k=10, release build, warm mmap scan buffer.
| L2 | ~12ms | ~47ms | **3.8× faster** |
| Cosine | ~13ms | ~42ms | **3.1× faster** |
| Dot | ~13ms | ~36ms | **2.8× faster** |
All three metrics use a single scan of the mmap float buffer; distance computation is SIMD-accelerated.
---
### INSERT
```sql
INSERT INTO table (col1, col2) VALUES (1, 'a')
INSERT INTO table VALUES (1, 'a', 3.14)
INSERT INTO table (col1, col2) VALUES (1, 'a'), (2, 'b'), (3, 'c')
```
### DDL (Data Definition Language)
ApexBase supports full SQL DDL operations.
#### CREATE TABLE
```sql
CREATE TABLE table_name
CREATE TABLE IF NOT EXISTS table_name
```
#### ALTER TABLE
```sql
-- Add column
ALTER TABLE table_name ADD COLUMN column_name DATA_TYPE
-- Rename column
ALTER TABLE table_name RENAME COLUMN old_name TO new_name
-- Drop column
ALTER TABLE table_name DROP COLUMN column_name
```
#### DROP TABLE
```sql
DROP TABLE table_name
DROP TABLE IF EXISTS table_name
```
#### Supported Data Types
| `STRING` | `VARCHAR`, `TEXT` | String/text data |
| `INT` | `INTEGER`, `INT32`, `INT64` | Integer numbers |
| `FLOAT` | `DOUBLE`, `FLOAT64` | Floating point numbers |
| `BOOL` | `BOOLEAN` | Boolean values |
| `FLOAT16_VECTOR` | `FLOAT16VECTOR`, `F16_VECTOR` | Half-precision float vector (SIMD-accelerated TopK) |
### Examples
```python
# Create table via SQL
client.execute("CREATE TABLE IF NOT EXISTS users")
# Add columns via SQL
client.execute("ALTER TABLE users ADD COLUMN name STRING")
client.execute("ALTER TABLE users ADD COLUMN age INT")
# Insert data via SQL
client.execute("INSERT INTO users (name, age) VALUES ('Alice', 30)")
results = client.execute("SELECT * FROM users WHERE age > 25")
# Modify schema via SQL
client.execute("ALTER TABLE users RENAME COLUMN name TO full_name")
client.execute("ALTER TABLE users DROP COLUMN age")
# Drop table via SQL
client.execute("DROP TABLE IF EXISTS users")
```
#### Multi-Statement SQL
You can execute multiple SQL statements in a single call by separating them with semicolons:
```python
# Execute multiple statements at once
client.execute("""
CREATE TABLE IF NOT EXISTS products;
ALTER TABLE products ADD COLUMN name STRING;
ALTER TABLE products ADD COLUMN price FLOAT;
INSERT INTO products (name, price) VALUES ('Laptop', 999.99)
""")
# Multiple INSERT statements
client.execute("""
INSERT INTO products (name, price) VALUES ('Mouse', 29.99);
INSERT INTO products (name, price) VALUES ('Keyboard', 79.99);
INSERT INTO products (name, price) VALUES ('Monitor', 299.99)
""")
# The result of the last statement is returned
results = client.execute("""
CREATE TABLE IF NOT EXISTS temp;
INSERT INTO temp (name) VALUES ('test');
SELECT * FROM temp
""")
```
**Multi-Statement SQL Rules:**
- Statements are separated by semicolons (`;`)
- Semicolons inside string literals are handled correctly
- Statements execute sequentially in order
- The result of the last SELECT statement is returned
- DDL statements (CREATE, ALTER, DROP, INSERT) return empty results
---