sqlrite (Python)

Python bindings for SQLRite — a small, embeddable SQLite clone written in Rust. Shape follows PEP 249 / the stdlib sqlite3 module, so most callers can pick it up without reading the docs.

Install

# From PyPI:
pip install sqlrite

# From source in a clone of the repo:
pip install maturin
cd sdk/python
maturin develop --release

Quick tour

import sqlrite

# File-backed or in-memory (use `":memory:"` to match sqlite3 convention).
conn = sqlrite.connect("foo.sqlrite")
# conn = sqlrite.connect(":memory:")

cur = conn.cursor()
cur.execute("CREATE TABLE users (id INTEGER PRIMARY KEY, name TEXT, age INTEGER)")
cur.execute("INSERT INTO users (name, age) VALUES ('alice', 30)")
cur.execute("INSERT INTO users (name, age) VALUES ('bob', 25)")

cur.execute("SELECT id, name, age FROM users")
for row in cur:
    print(row)  # (1, 'alice', 30), then (2, 'bob', 25)

# Column metadata (PEP 249 `.description`):
for col in cur.description:
    print(col[0])  # 'id', 'name', 'age'

conn.close()

Transactions

with sqlrite.connect("foo.sqlrite") as conn:
    cur = conn.cursor()
    cur.execute("BEGIN")
    cur.execute("INSERT INTO users (name) VALUES ('carol')")
    if looks_good:
        conn.commit()
    else:
        conn.rollback()
# The context manager automatically commits on clean exit
# and rolls back on exception, then closes the connection.

Read-only access

conn = sqlrite.connect_read_only("foo.sqlrite")
# Any write raises sqlrite.SQLRiteError. Multiple read-only
# connections on the same file coexist (shared OS lock).

Vector columns + KNN (Phase 7a–7d)

The engine ships with a fixed-dimension VECTOR(N) storage class and three distance functions (vec_distance_l2, vec_distance_cosine, vec_distance_dot). Plain cursor.execute(...) is all you need from Python — values come back as Python lists of floats.

cur.execute("CREATE TABLE docs (id INTEGER PRIMARY KEY, embedding VECTOR(384))")
cur.execute("INSERT INTO docs (id, embedding) VALUES (1, [0.1, 0.2, ..., 0.0])")  # 384 floats
cur.execute("""
    SELECT id FROM docs
     ORDER BY vec_distance_l2(embedding, [0.1, 0.2, ..., 0.0])
     LIMIT 10
""")
for row in cur:
    print(row)

For larger collections, build an HNSW index — the executor will use it automatically when the WHERE/ORDER BY shape matches:

cur.execute("CREATE INDEX idx_docs_emb ON docs USING hnsw (embedding)")

JSON columns (Phase 7e)

JSON (and JSONB as an alias) columns store text, validated at INSERT/UPDATE time. Read with json_extract / json_type / json_array_length / json_object_keys. Path subset: $, .key, [N], chained.

cur.execute("CREATE TABLE events (id INTEGER PRIMARY KEY, payload JSON)")
cur.execute(
    "INSERT INTO events (payload) VALUES "
    "('{\"user\": {\"name\": \"alice\"}, \"score\": 42}')"
)
cur.execute(
    "SELECT json_extract(payload, '$.user.name'), json_type(payload, '$.score') FROM events"
)
print(cur.fetchall())  # [('alice', 'integer')]

json_object_keys returns a JSON-array text rather than a table-valued result (set-returning functions aren't supported yet).

Natural-language → SQL (Phase 7g.4)

Connection.ask(question) generates SQL from a natural-language question via the configured LLM provider (Anthropic by default). Connection.ask_run(question) is the convenience that calls ask() then immediately executes the generated SQL.

import sqlrite

conn = sqlrite.connect("foo.sqlrite")
conn.execute("CREATE TABLE users (id INTEGER PRIMARY KEY, name TEXT, age INTEGER)")
conn.execute("INSERT INTO users (name, age) VALUES ('alice', 30)")

# Reads SQLRITE_LLM_API_KEY from the environment.
resp = conn.ask("How many users are over 30?")
print(resp.sql)            # "SELECT COUNT(*) FROM users WHERE age > 30"
print(resp.explanation)    # "Counts users over the age threshold."

# Caller decides whether to run the SQL — ask() does NOT auto-execute.
cur = conn.cursor()
cur.execute(resp.sql)
print(cur.fetchall())      # [(1,)]

# Or one-shot:
rows = conn.ask_run("How many users are over 30?").fetchall()

Configuration

Three precedence layers — explicit-wins:

1. Per-call config: conn.ask("...", sqlrite.AskConfig(api_key="..."))
2. Per-connection config: conn.set_ask_config(cfg) — applies to all subsequent ask() / ask_run() calls on this connection
3. Environment vars (zero-config fallback): SQLRITE_LLM_PROVIDER / _API_KEY / _MODEL / _MAX_TOKENS / _CACHE_TTL

# Path 1: env (zero config) — same env vars as REPL/Desktop
resp = conn.ask("how many users?")

# Path 2: explicit per-call (highest precedence)
cfg = sqlrite.AskConfig(
    api_key="sk-ant-...",
    model="claude-haiku-4-5",
    max_tokens=512,
    cache_ttl="1h",        # "5m" (default) | "1h" | "off"
)
resp = conn.ask("how many users?", cfg)

# Path 3: per-connection (set once, reuse)
conn.set_ask_config(cfg)
resp = conn.ask("how many users?")     # uses cfg
resp = conn.ask("count by age")        # uses cfg

# Or build from env explicitly:
cfg = sqlrite.AskConfig.from_env()

Defaults

provider="anthropic", model="claude-sonnet-4-6", max_tokens=1024, cache_ttl="5m". The schema dump goes inside an Anthropic prompt-cache breakpoint — repeat asks against the same DB hit the cache (verify via resp.usage.cache_read_input_tokens).

Errors

Missing API key → sqlrite.SQLRiteError("missing API key (set SQLRITE_LLM_API_KEY ...)"). API errors (4xx/5xx) bubble up as SQLRiteError with the status code + Anthropic's structured error message. ask_run() on an empty SQL response (model declined to generate) raises with the model's explanation rather than executing the empty string.

What AskResponse carries

resp.sql              # str — generated SQL, ready to execute (or '' if model declined)
resp.explanation      # str — one-sentence rationale (may be empty)
resp.usage.input_tokens
resp.usage.output_tokens
resp.usage.cache_creation_input_tokens
resp.usage.cache_read_input_tokens

AskConfig.__repr__ and AskResponse.__repr__ deliberately don't include the API key — printing config in logs won't leak it.

API surface

Parameter binding

execute(sql, params) accepts None and empty tuples/lists for DB-API compatibility, but any non-empty params raises TypeError with a clear message — the underlying engine doesn't support prepared-statement parameter binding yet (deferred to Phase 5a.2). For now, inline values into the SQL (with manual escaping).

Running the tests

maturin develop
python -m pytest tests/

How this ships

• PyO3 (abi3-py38) for the Rust-Python boundary — one wheel works on every CPython 3.8+ release, no per-version rebuild.
• maturin as the build backend, emitting standard .whl files that pip can install directly.
• Phase 6f's CI publishes abi3-py38 wheels to PyPI for manylinux x86_64/aarch64, macOS aarch64, and Windows x86_64 (plus an sdist) on every release. OIDC trusted publishing — no long-lived PyPI token in the repo.

Sibling products

This SDK is for when your code drives the database. If you want an LLM agent to drive a SQLRite database directly, install the sqlrite-mcp Model Context Protocol server (cargo install sqlrite-mcp) and wire it into Claude Code / Cursor / mcp-inspector / any MCP-aware client. Same engine underneath.

Status

Phase 5c MVP: ✅ — basic CRUD, transactions, context managers, read-only mode, iteration. Parameter binding, CursorRow namedtuples, and type converters (datetime, Decimal) are natural follow-ups.