Zova Python Bindings

This package contains the source-first Python bindings for Zova.

It is a PyO3/maturin extension backed by the safe Rust zova binding. It does not wrap the C ABI directly with ctypes or cffi. The native build still uses Zova's C ABI underneath through the Rust zova-sys crate, so Python gets the same records/objects/vectors foundation without reimplementing the ABI ownership rules.

Contents

1. How It Fits
2. Install
3. Local Development
4. What It Covers
5. Savepoints
6. Operational Safety
7. Extensions
8. App Events
9. Objects
10. Vectors
11. Graphs
12. Bound Stores
13. SQL-Native Vector Search

How It Fits

Python users import zova. The extension is built with PyO3 and reuses the safe Rust binding, which in turn links Zova's C ABI.

flowchart LR
    App["Python app"]
    PyPkg["zova Python package"]
    PyO3["PyO3 extension"]
    Rust["Rust zova crate"]
    CABI["libzova_c.a"]
    File["local .zova file"]

    App --> PyPkg
    PyPkg --> PyO3
    PyO3 --> Rust
    Rust --> CABI
    CABI --> File

Install

After the Python package is published to PyPI:

python -m pip install zova

The Python package is backed by a PyO3 extension built through maturin and Cargo. Users need Python 3.13 or newer; CPython 3.13 and 3.14 are the tested runtime matrix. Linux wheels target glibc 2.28 or newer. Published stable-ABI wheels do not require Rust, Cargo, Zig, or a local C compiler. Zova does not publish a Python source distribution, so an unsupported platform or interpreter fails without attempting a native build.

The release workflow builds Linux/macOS x86_64/arm64 and Windows x86_64 stable-ABI wheels.

The current release uses .zova format 12 and does not migrate older format databases in place. Format-9, format-10 and format-11 databases can be probed and migrated forward with the module-level probe_format(path) and migrate_database(source, destination, *, verify=True) functions: migration is explicit, copy-forward, publishes a separately validated format-12 destination, and never mutates the source.

Opaque-key graph, edge-payload, topology-scan, prepared-build, and generic fresh-build session APIs are supported low-level C/raw zova-sys surfaces. The Python package deliberately exposes the established graph CRUD and traversal API instead of those specialized publication surfaces.

Local Development

From bindings/python:

uv run --isolated --with maturin --with pytest maturin develop
uv run --isolated --with pytest python -m pytest

The native build uses maturin, Cargo, and the Rust zova crate. Users do not need to locate a shared C library manually.

The Python API is part of the stable 1.x contract described in ../../API_STABILITY.md.

What It Covers

The Python package exposes database lifecycle, conversion, prepared SQL statements, transactions, explicit vacuum, backup/compact/restore, objects, streaming object writes, vectors, SQL-native vector search, graphs, bundled extension lifecycle, same-process app events, context managers, and Zova status exceptions.

One Python Database object owns one native handle. The native C ABI serializes calls on that handle, so one handle is safe but not parallel. Open additional database handles when an application needs independent concurrent connections; SQLite locking rules still apply across handles. PyO3 classes remain unsendable in this release even though the native C ABI serializes its own calls.

Use Database.open(path, read_only=True) for read-only handles, and Database.set_busy_timeout(milliseconds) when an application wants SQLite to wait briefly on cross-handle contention. No nonzero timeout is installed by default.

Use Database.last_insert_rowid(), Database.changes(), Database.total_changes(), and Statement.column_name(index) for normal application SQL record helpers. They do not expose or stabilize Zova's private _zova_* tables.

Savepoints

Use explicit savepoints for partial rollback inside one database connection:

with zova.Database.open("app.zova") as db:
    db.begin_immediate()
    db.savepoint("attach_file")
    db.exec("insert into attachments(filename) values ('draft.txt')")
    db.rollback_to_savepoint("attach_file")
    db.release_savepoint("attach_file")
    db.commit()

Savepoint names are strict ASCII identifiers: 1-64 bytes, first byte [A-Za-z_], remaining bytes [A-Za-z0-9_], and no case-insensitive _zova_ prefix. rollback_to_savepoint() keeps the savepoint active; release_savepoint() removes it. An inner released savepoint can still be undone by rolling back an outer transaction or savepoint.

Use savepoint_context() when you want rollback cleanup tied to a with block:

with db.savepoint_context("attach_file") as scoped_db:
    scoped_db.exec("insert into attachments(filename) values ('draft.txt')")

On normal exit the context releases the savepoint. On exception it rolls back, releases, and then re-raises the original exception when cleanup succeeds.

Operational Safety

Use backup_to() for a faithful snapshot, compact_to() for a space-reclaiming copy, and restore_backup() to copy a backup into a new destination file. Destinations must be .zova paths and are never overwritten.

with zova.Database.open("app.zova") as db:
    db.backup_to("app.backup.zova")
    db.compact_to("app.compact.zova")

zova.restore_backup("app.backup.zova", "app.restored.zova")

By default, each operation verifies the destination after copying. Pass verify=False only when you will verify separately, for example with zova check --deep.

Diagnostic recovery commands such as zova doctor, zova salvage --dry-run, and zova salvage <source> <destination> are CLI-first. In v0.25, CLI salvage copies valid core data and uses extension salvage hooks for extension-owned storage. The Python package does not expose typed doctor/salvage report APIs yet, and library code should not parse the human text output as a stable binding contract.

Extensions

Python exposes lifecycle methods for extensions already present in the current process registry. The default registry includes bundled extensions such as trgm, so Python applications can install, list, check, and drop trgm directly. From 1.2.0, wheels also accept explicit trusted native bundles:

db = zova.Database.create("app.zova", extension_bundles=["./search.zovaext"])
db.install_extension("search")
db.close()
db = zova.Database.open("app.zova", read_only=True,
                        extension_bundles=["./search.zovaext"])

Both constructors accept trust_store_path; omit it for Zova's default store. First run zova extension trust ./search.zovaext. Opening never grants trust or loads paths from the database. Native code and dependencies must match the host. Owned statements retain their connection and libraries after db.close(). Published 1.1.0 wheels do not include this support. Python does not wrap extension authoring, application scalar callbacks or explicit data upgrades. See the capability matrix.

See ../../docs/extensions.md for the current host contract and trust model. A fuller records/objects/vectors/graphs example lives in examples/extensions.py.

with zova.Database.open("app.zova") as db:
    db.install_extension("trgm")
    info = db.extension_info("trgm")
    assert info.storage_prefix == "_zova_ext_trgm_"
    db.check_extensions()

    with db.prepare(
        "select zova_trgm_put('messages', ?1, 'record', 'messages', ?2, ?3)"
    ) as stmt:
        stmt.bind_text(1, "message:123")
        stmt.bind_text(2, "123")
        stmt.bind_text(3, "attachment upload failed")
        assert stmt.step() == zova.Step.ROW

    with db.prepare(
        """
        select document_id, score
        from zova_trgm_search
        where index_name = 'messages'
          and query = ?1
          and "limit" = 10
        order by rank
        """
    ) as stmt:
        stmt.bind_text(1, "attachement failed")

App Events

Use listen / notify for same-process storage workflow notifications. They are explicit, in-memory, local to one open Database handle, and delivered only after the surrounding Zova transaction commits. Rollback discards pending notifications.

with zova.Database.open("app.zova") as db:
    with db.listen("message:123:attachments") as sub:
        db.begin_immediate()
        db.exec("insert into attachments(message_id, name) values (123, 'photo.jpg')")
        db.notify("message:123:attachments", "changed")
        db.commit()

        note = sub.try_receive()
        if note is not None:
            print(note.channel, note.payload)

SQL zova_notify(...) follows the same transaction rules when the surrounding transaction/savepoint was opened through Zova helpers; raw SQL transaction scopes are rejected because Zova cannot track their notification lifetime.

Event delivery is queue-only in v0.18: no callbacks, no blocking receive, no cross-process delivery, no replay after restart, and no automatic logging of SQL, object, vector, or graph mutations. Each subscription queue holds 1024 notifications and drops the oldest entries on overflow; the next received notification reports how many were dropped before it.

Objects

The Python binding exposes Zova objects as content-addressed byte values while keeping application metadata in normal SQL tables.

import zova

with zova.Database.create("app.zova") as db:
    db.exec(
        "create table attachments("
        "id integer primary key, "
        "filename text not null, "
        "object_id blob not null)"
    )

    object_id = db.put_object(b"hello from Zova")

    with db.prepare("insert into attachments(filename, object_id) values (?1, ?2)") as stmt:
        stmt.bind_text(1, "hello.txt")
        stmt.bind_blob(2, bytes(object_id))
        stmt.step()

    assert db.read_object_range(object_id, 0, 5) == b"hello"

For large inputs, use ObjectWriter so the full object does not have to be held in memory by the caller:

with db.object_writer() as writer:
    writer.write(b"chunk one")
    writer.write(b"chunk two")
    object_id = writer.finish()

If a writer leaves the context without finish(), it is cancelled and any unreferenced chunks written by that writer are cleaned up. Writer operations follow Zova's object transaction policy in single-file mode. When an object store is bound, object writes participate in the same Zova transaction/savepoint scope through SQLite ATTACH.

Loose chunks and assembly are also exposed for receive-side workflows: applications track transfer state in their own SQL tables, call put_object_chunk() for verified chunks, then call assemble_object_from_chunks() when the manifest is complete.

Vectors

The Python binding exposes Zova vectors with the same model as the Rust binding: Zova stores numeric vectors in named collections, while application metadata stays in normal SQL tables.

import zova

with zova.Database.create("vectors.zova") as db:
    db.exec(
        "create table chunks("
        "id text primary key, "
        "document_id text not null, "
        "text text not null, "
        "vector_id text not null)"
    )

    db.create_vector_collection(
        "chunks",
        zova.VectorCollectionOptions(2, zova.VectorMetric.L2, zova.VectorElementType.F32),
    )
    db.put_vectors(
        "chunks",
        [
            zova.VectorInput("chunk:1", zova.VectorElementType.F32, [0.0, 0.0]),
            zova.VectorInput("chunk:2", zova.VectorElementType.F32, [1.0, 0.0]),
        ],
    )

    db.exec(
        "insert into chunks(id, document_id, text, vector_id) values "
        "('c1', 'doc-a', 'first chunk', 'chunk:1'), "
        "('c2', 'doc-a', 'near chunk', 'chunk:2')"
    )

    results = db.search_vectors_in(
        "chunks",
        zova.VectorElementType.F32,
        [0.0, 0.0],
        ["chunk:1", "chunk:2"],
        2,
    )
    for result in results:
        with db.prepare("select text from chunks where vector_id = ?1") as stmt:
            stmt.bind_text(1, result.id)
            stmt.step()
            print(result.id, result.distance, stmt.column_text(0))

Vector APIs are typed by default. F32 uses Python floats, F16 values are raw IEEE 754 binary16 bits carried as integer uint16 bit patterns, and I8 values are raw signed byte values with no quantization metadata.

db.create_vector_collection(
    "scores_i8",
    zova.VectorCollectionOptions(2, zova.VectorMetric.L2, zova.VectorElementType.I8),
)
db.put_vector("scores_i8", "near", zova.VectorElementType.I8, [1, -1])

db.create_vector_collection(
    "halves",
    zova.VectorCollectionOptions(2, zova.VectorMetric.L2, zova.VectorElementType.F16),
)
db.put_vector("halves", "one", zova.VectorElementType.F16, [0x3C00, 0x0000])
assert db.get_vector("halves", "one").values == [0x3C00, 0x0000]

Search is exact and lower distance is better. Candidate-filtered searches skip missing ids and deduplicate duplicate candidates. Search-by-id excludes the source vector. Threshold variants are inclusive, and dot-product thresholds may be negative because dot distance is -dot_product.

Deleting a vector collection removes Zova's private vector rows only. User SQL metadata rows that reference vector ids are application-owned and remain in place.

Graphs

Graphs store relationship topology while application metadata stays in SQL tables, objects, and vectors. Apps provide stable node IDs and can point nodes at records, objects, object chunks, vectors, entities, facts, concepts, or external references.

db.create_graph(zova.DEFAULT_GRAPH_NAME)
db.put_graph_node(
    zova.GraphNodeInput(
        zova.DEFAULT_GRAPH_NAME,
        "message:1",
        "message",
        zova.GraphTargetType.RECORD,
        "messages",
        "1",
    )
)
db.put_graph_node(
    zova.GraphNodeInput(
        zova.DEFAULT_GRAPH_NAME,
        "entity:zova",
        "entity",
        zova.GraphTargetType.ENTITY,
        None,
        "zova",
    )
)
db.put_graph_edge(
    zova.GraphEdgeInput(
        zova.DEFAULT_GRAPH_NAME,
        "message:1",
        "mentions",
        "entity:zova",
    )
)

Use graph_neighbors() for one-hop expansion and graph_walk() for bounded directed walks. Zova validates object, chunk, and vector targets it owns, but arbitrary SQL row existence remains the application's job. Node IDs may contain sensitive app identifiers, so choose export-safe IDs when files may leave the app.

SQL-native graph helpers are available through ordinary prepared statements:

with db.prepare(
    "select m.body "
    "from zova_graph_neighbors as g "
    "join messages as m on m.graph_node_id = g.node_id "
    "where g.graph_name = 'default' "
    "and g.source_node_id = 'message:1' "
    'and g."limit" = 20 '
    "order by g.rank"
) as stmt:
    ...

Use zova_graph_neighbors for one-hop joins and zova_graph_walk for bounded directed walks with depth, predecessor_node_id, and edge_type columns.

Bound Stores

In v0.25, a .zova file may be bound to one object store, one vector store, and one graph store through the native Zig API or CLI. The Python object, vector, and graph methods above transparently use those stores after Database.open. Store create/bind/unbind/split management is not exposed as a Python API yet.

SQL-Native Vector Search

Zova registers SQL vector functions and the zova_vector_search virtual table on Zova database connections. Bind query vectors as little-endian f32 blobs with encode_f32_le():

query = zova.encode_f32_le([0.0, 0.0])

with db.prepare(
    "select c.id, zova_vector_distance('chunks', c.vector_id, ?1) as distance "
    "from chunks as c "
    "where c.document_id = 'doc-a' "
    "order by distance "
    "limit 10"
) as stmt:
    stmt.bind_blob(1, query)

Row-to-row distances use zova_vector_distance_by_id(collection, vector_id, source_vector_id). Collection-wide SQL search uses zova_vector_search:

with db.prepare(
    "select c.text, s.distance "
    "from zova_vector_search as s "
    "join chunks as c on c.vector_id = s.vector_id "
    "where s.collection = 'chunks' "
    "and s.query_vector = ?1 "
    "and s.top_k = 10 "
    "order by s.rank"
) as stmt:
    stmt.bind_blob(1, query)

Python's built-in sqlite3 module opens an ordinary SQLite connection and does not automatically register Zova's SQL functions or virtual table. Use zova.Database for SQL-native vector search in this binding. A future SQLite loadable extension may make the SQL surface available to external SQLite connections.