Build a Vector Extension for Postgres - Vector Type

[SteveLauC]

In this post, we are going to implement the vector type for Postgres:

CREATE TABLE items (v vector(3));

Postgres Extension Structure and pgrx Wrappers

Before we implement it, let's take a look at the typical extension structure, and how pgrx simplifies it for us.

A typical Postgres extension can be roughly split into 3 layers:

1. The implementation, which is usually done in a low-level language like C.
2. Higher level SQL statements that glue the implementation to Postgres.
3. A control file that specifies a few basic properties of the extension.

If you take a look at the source code of pgvector, this 3-layer structure is quite obvious, the src directory is for C code, the sql directory contains higher SQL glues, and there is also a control file. So how does pgrx make extension building easier?

1. It wraps the Postgres C APIs in Rust

   As we said, even though we build extensions in Rust, Postgres's APIs are still C, pgrx tries to wrap them in Rust so we don't need to bother with C.

2. SQL glues are generated using Rust macros, if possible

   Later we will see that pgrx could generate the SQL glues for us automatically.

3. pgrx generates the control file for us

CREATE TYPE vector

Let's define our Vector type, using std::vec::Vec looks pretty straightforward, and since vector needs to store floats, we use f64 here:

struct Vector {
    value: Vec<f64>
}

Then what?

The SQL statement used to create new types is CREATE TYPE ..., from its documentation, we would know that the vector type we are implementing is a base type, to create a base type, support functions input_function and output_function are required. And since it needs to take a dimension argument (vector(DIMENSION)), which is implemented using modifer, functions type_modifier_input_function and type_modifier_output_function are also needed. So we need to implement these 4 functions for our Vector type.

input_function

To quote the documentation,

The input_function converts the type's external textual representation to the internal representation used by the operators and functions defined for the type.

The input function can be declared as taking one argument of type cstring, or as taking three arguments of types cstring, oid, integer. The first argument is the input text as a C string, the second argument is the type's own OID (except for array types, which instead receive their element type's OID), and the third is the typmod of the destination column, if known (-1 will be passed if not). The input function must return a value of the data type itself.

Ok, from the documentation, this input_function is used for deserialization, serde is the most popular de/serialization library in Rust, so let's use it. For the arguments, since the vector needs a type modifier, we need it to take 3 arguments. Here is what our input_function looks like:

#[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
fn vector_input(
    input: &CStr,
    _oid: pg_sys::Oid,
    type_modifier: i32,
) -> Vector {
    let value = match serde_json::from_str::<Vec<f64>>(
        input.to_str().expect("expect input to be UTF-8 encoded"),
    ) {
        Ok(v) => v,
        Err(e) => {
            pgrx::error!("failed to deserialize the input string due to error {}", e)
        }
    };
    let dimension = match u16::try_from(value.len()) {
        Ok(d) => d,
        Err(_) => {
            pgrx::error!("this vector's dimension [{}] is too large", value.len());
        }
    };
    
    // cast should be safe as dimension should be a positive
    let expected_dimension = match u16::try_from(type_modifier) {
        Ok(d) => d,
        Err(_) => {
            panic!("failed to cast stored dimension [{}] to u16", type_modifier);
        }
    };

    // check the dimension 
    if dimension != expected_dimension {
        pgrx::error!(
            "mismatched dimension, expected {}, found {}",
            expected_dimension,
            dimension
        );
    }

    Vector { value }
}

That's a bunch of stuff, let's go through it piece by piece.

#[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]

If you mark a function with pg_extern, then pgrx will automatically generate SQL like CREATE FUNCTION <your function> for you, immutable, strict, parallel_safe are the attributes that you think your function has, they correspond to the attributes listed in the CREATE FUNCTION doc. Because this Rust macro is used to generate SQL, and SQLs could depend on each other, this requires = [ "shell_type" ] is used to clarify this dependency relationship.

shell_type is the name of another SQL snippet that defines the shell type, what is shell type? It behaves like a placeholder so that we can have a vector type to use before we fully implement it. The SQL generated by this #[pg_extern] macro would be:

CREATE FUNCTION "vector_input"(
	"input" cstring,
	"_oid" oid, 
	"type_modifier" INT 
) RETURNS vector

As you can see, this function RETURNS vector, but how can we have a vector type before we implement these 4 required functions?

Shell type is exactly for this! We can define a shell type (A dummy type, no function needs to be provided), and let our functions depend on it:

pgrx won't define this shell type for us, we need to manually do this in SQL, here is what it looks like:

extension_sql!(
    r#"CREATE TYPE vector; -- shell type"#,
    name = "shell_type"
);

The extension_sql!() macro allows us to write SQL in Rust code, then pgrx will include it in the generated SQL script. The name = "shell_type" specifies this SQL snippet's identifier and can be used to refer to it. Our vector_input() function depends on this shell type, so it requires = [ "shell_type" ].

fn vector_input(
    input: &CStr,
    _oid: pg_sys::Oid,
    type_modifier: i32,
) -> Vector {

The input argument is a string representing our vector input text, _oid is prefixed with _ because we do not need it. The type_modifier argument is of type i32, this is how type modifier gets stored in Postgres. We will see it again when we implement the type modifier input/output functions.

let value = match serde_json::from_str::<Vec<f64>>(
    input.to_str().expect("expect input to be UTF-8 encoded"),
) {
    Ok(v) => v,
    Err(e) => {
        pgrx::error!("failed to deserialize the input string due to error {}", e)
    }
};

Then we convert input to a UTF-8 encoded &str and pass it to serde_json::from_str(). The input text should be UTF-8 encoded, so we should be safe. If any error happens during deserialization, just error out with pgrx::error!(), which will log at the error level and terminate the current transaction.

let dimension = match u16::try_from(value.len()) {
    Ok(d) => d,
    Err(_) => {
        pgrx::error!("this vector's dimension [{}] is too large", value.len());
    }
};

// cast should be safe as dimension should be a positive
let expected_dimension = match u16::try_from(type_modifier) {
    Ok(d) => d,
    Err(_) => {
        panic!("failed to cast stored dimension [{}] to u16", type_modifier);
    }
};

The max dimension supported by us is u16::MAX, we do this simply because this is what pgvector does.

// check the dimension 
if dimension != expected_dimension {
    pgrx::error!(
        "mismatched dimension, expected {}, found {}",
        expected_dimension,
        dimension
    );
}

Vector { value }

Lastly, we check if the input vector has the expected dimension, and error out if not. Otherwise, we return the parsed vector.

output_function

The output_function performs the reverse action, it serializes the given vector to a string. Here is our implementation:

#[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
fn vector_output(value: Vector) -> CString {
    let value_serialized_string = serde_json::to_string(&value).unwrap();
    CString::new(value_serialized_string).expect("there should be no NUL in the middle")
}

We just serialize the Vec<f64> and return it in a CString, quite simple.

type_modifier_input_function

The type_modifier_input_function should parse the input modifier(s), check the parsed modifier(s), and if valid, encode it(them) into an integer, which is how Postgres stores type modifier.

A type can accept multiple type modifiers separated by ,, this is the reason why we see an array here.

#[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
fn vector_modifier_input(list: pgrx::datum::Array<&CStr>) -> i32 {
    if list.len() != 1 {
        pgrx::error!("too many modifiers, expect 1")
    }

    let modifier = list
        .get(0)
        .expect("should be Some as len = 1")
        .expect("type modifier cannot be NULL");
    let Ok(dimension) = modifier
        .to_str()
        .expect("expect type modifiers to be UTF-8 encoded")
        .parse::<u16>()
    else {
        pgrx::error!("invalid dimension value, expect a number in range [1, 65535]")
    };

    dimension as i32
}

The implementation is straightforward, the vector type takes only 1 modifier, so we verify if the array length is 1. And if yes, try to parse it to a u16, if no error happens, return it as an i32 so that Postgres can store it.

type_modifier_output_function

#[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
fn vector_modifier_output(type_modifier: i32) -> CString {
    CString::new(format!("({})", type_modifier)).expect("no NUL in the middle")
}

type_modifier_output_function implementation is also simple, just return a string in format (type_modifier).

All the required functions are implemented, now we can finally CREATE it!

// create the actual type, specifying the input and output functions
extension_sql!(
    r#"
CREATE TYPE vector (
    INPUT = vector_input,
    OUTPUT = vector_output,
    TYPMOD_IN = vector_modifier_input,
    TYPMOD_OUT = vector_modifier_output,
    STORAGE = external 
);
"#,
    name = "concrete_type",
    creates = [Type(Vector)],
    requires = [
        "shell_type",
        vector_input,
        vector_output,
        vector_modifier_input,
        vector_modifier_output
    ]
);

You may be unfamiliar with the STORAGE = external argument. Let me give a brief introduction here. The default storage engine used by Postgres is heap, with this engine, a table's the on-disk file is split into pages, whose size defaults to 8192 bytes, Postgres hopes that at least 4 rows can be stored in a page, if some columns are too big so that 4 rows cannot be accommodated, Postgres moves them to an external table, which is call a TOAST table. This STORAGE argument controls Postgres's behavior when moving columns into TOAST tables. Setting it to external means this column can be moved to the TOAST table if it is too huge.

Our first launch!

Looks like everything is ready, let's do our first launch!

$ cargo pgrx run
error[E0277]: the trait bound `for<'a> fn(&'a CStr, Oid, i32) -> Vector: FunctionMetadata<_>` is not satisfied
  --> pg_extension/src/vector_type.rs:87:1
   |
86 | #[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
   | --------------------------------------------------------------------------- in this procedural macro expansion
87 | fn vector_input(input: &CStr, _oid: pg_sys::Oid, type_modifier: i32) -> Vector {
   | ^^ the trait `FunctionMetadata<_>` is not implemented for `for<'a> fn(&'a CStr, Oid, i32) -> Vector`
   |
   = help: the following other types implement trait `FunctionMetadata<A>`:
             `unsafe fn() -> R` implements `FunctionMetadata<()>`
             `unsafe fn(T0) -> R` implements `FunctionMetadata<(T0,)>`
             `unsafe fn(T0, T1) -> R` implements `FunctionMetadata<(T0, T1)>`
             `unsafe fn(T0, T1, T2) -> R` implements `FunctionMetadata<(T0, T1, T2)>`
             `unsafe fn(T0, T1, T2, T3) -> R` implements `FunctionMetadata<(T0, T1, T2, T3)>`
             `unsafe fn(T0, T1, T2, T3, T4) -> R` implements `FunctionMetadata<(T0, T1, T2, T3, T4)>`
             `unsafe fn(T0, T1, T2, T3, T4, T5) -> R` implements `FunctionMetadata<(T0, T1, T2, T3, T4, T5)>`
             `unsafe fn(T0, T1, T2, T3, T4, T5, T6) -> R` implements `FunctionMetadata<(T0, T1, T2, T3, T4, T5, T6)>`
           and 25 others
   = note: this error originates in the attribute macro `pg_extern` (in Nightly builds, run with -Z macro-backtrace for more info)

error[E0599]: the function or associated item `entity` exists for struct `Vector`, but its trait bounds were not satisfied
  --> pg_extension/src/vector_type.rs:86:1
   |
22 | pub(crate) struct Vector {
   | ------------------------ function or associated item `entity` not found for this struct because it doesn't satisfy `Vector: SqlTranslatable`
...
86 | #[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
   | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ function or associated item cannot be called on `Vector` due to unsatisfied trait bounds
   |
   = note: the following trait bounds were not satisfied:
           `Vector: SqlTranslatable`
           which is required by `&Vector: SqlTranslatable`
note: the trait `SqlTranslatable` must be implemented
  --> $HOME/.cargo/registry/src/index.crates.io-6f17d22bba15001f/pgrx-sql-entity-graph-0.12.9/src/metadata/sql_translatable.rs:73:1
   |
73 | pub unsafe trait SqlTranslatable {
   | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
   = help: items from traits can only be used if the trait is implemented and in scope
   = note: the following traits define an item `entity`, perhaps you need to implement one of them:
           candidate #1: `FunctionMetadata`
           candidate #2: `SqlTranslatable`
   = note: this error originates in the attribute macro `pg_extern` (in Nightly builds, run with -Z macro-backtrace for more info)

error[E0277]: the trait bound `Vector: RetAbi` is not satisfied
  --> pg_extension/src/vector_type.rs:87:73
   |
87 | fn vector_input(input: &CStr, _oid: pg_sys::Oid, type_modifier: i32) -> Vector {
   |                                                                         ^^^^^^ the trait `BoxRet` is not implemented for `Vector`, which is required by `Vector: RetAbi`
   |
   = help: the following other types implement trait `BoxRet`:
             &'a CStr
             &'a [u8]
             &'a str
             ()
             AnyArray
             AnyElement
             AnyNumeric
             BOX
           and 33 others
   = note: required for `Vector` to implement `RetAbi`

error[E0277]: the trait bound `Vector: RetAbi` is not satisfied
   --> pg_extension/src/vector_type.rs:87:80
    |
86  |   #[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
    |   --------------------------------------------------------------------------- in this procedural macro expansion
87  |   fn vector_input(input: &CStr, _oid: pg_sys::Oid, type_modifier: i32) -> Vector {
    |  ________________________________________________________________________________^
88  | |     let value = match serde_json::from_str::<Vec<f64>>(
89  | |         input.to_str().expect("expect input to be UTF-8 encoded"),
90  | |     ) {
...   |
110 | |     Vector { value }
111 | | }
    | |_^ the trait `BoxRet` is not implemented for `Vector`, which is required by `Vector: RetAbi`
    |
    = help: the following other types implement trait `BoxRet`:
              &'a CStr
              &'a [u8]
              &'a str
              ()
              AnyArray
              AnyElement
              AnyNumeric
              BOX
            and 33 others
    = note: required for `Vector` to implement `RetAbi`
    = note: this error originates in the attribute macro `pg_extern` (in Nightly builds, run with -Z macro-backtrace for more info)

error[E0277]: the trait bound `fn(Vector) -> CString: FunctionMetadata<_>` is not satisfied
   --> pg_extension/src/vector_type.rs:114:1
    |
113 | #[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
    | --------------------------------------------------------------------------- in this procedural macro expansion
114 | fn vector_output(value: Vector) -> CString {
    | ^^ the trait `FunctionMetadata<_>` is not implemented for `fn(Vector) -> CString`
    |
    = help: the following other types implement trait `FunctionMetadata<A>`:
              `unsafe fn() -> R` implements `FunctionMetadata<()>`
              `unsafe fn(T0) -> R` implements `FunctionMetadata<(T0,)>`
              `unsafe fn(T0, T1) -> R` implements `FunctionMetadata<(T0, T1)>`
              `unsafe fn(T0, T1, T2) -> R` implements `FunctionMetadata<(T0, T1, T2)>`
              `unsafe fn(T0, T1, T2, T3) -> R` implements `FunctionMetadata<(T0, T1, T2, T3)>`
              `unsafe fn(T0, T1, T2, T3, T4) -> R` implements `FunctionMetadata<(T0, T1, T2, T3, T4)>`
              `unsafe fn(T0, T1, T2, T3, T4, T5) -> R` implements `FunctionMetadata<(T0, T1, T2, T3, T4, T5)>`
              `unsafe fn(T0, T1, T2, T3, T4, T5, T6) -> R` implements `FunctionMetadata<(T0, T1, T2, T3, T4, T5, T6)>`
            and 25 others
    = note: this error originates in the attribute macro `pg_extern` (in Nightly builds, run with -Z macro-backtrace for more info)

error[E0599]: the function or associated item `entity` exists for struct `Vector`, but its trait bounds were not satisfied
   --> pg_extension/src/vector_type.rs:113:1
    |
22  | pub(crate) struct Vector {
    | ------------------------ function or associated item `entity` not found for this struct because it doesn't satisfy `Vector: SqlTranslatable`
...
113 | #[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
    | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ function or associated item cannot be called on `Vector` due to unsatisfied trait bounds
    |
    = note: the following trait bounds were not satisfied:
            `Vector: SqlTranslatable`
            which is required by `&Vector: SqlTranslatable`
note: the trait `SqlTranslatable` must be implemented
   --> $HOME/.cargo/registry/src/index.crates.io-6f17d22bba15001f/pgrx-sql-entity-graph-0.12.9/src/metadata/sql_translatable.rs:73:1
    |
73  | pub unsafe trait SqlTranslatable {
    | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    = help: items from traits can only be used if the trait is implemented and in scope
    = note: the following traits define an item `entity`, perhaps you need to implement one of them:
            candidate #1: `FunctionMetadata`
            candidate #2: `SqlTranslatable`
    = note: this error originates in the attribute macro `pg_extern` (in Nightly builds, run with -Z macro-backtrace for more info)

error[E0277]: the trait bound `Vector: ArgAbi<'_>` is not satisfied
   --> pg_extension/src/vector_type.rs:114:18
    |
113 | #[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
    | --------------------------------------------------------------------------- in this procedural macro expansion
114 | fn vector_output(value: Vector) -> CString {
    |                  ^^^^^ the trait `ArgAbi<'_>` is not implemented for `Vector`
    |
    = help: the following other types implement trait `ArgAbi<'fcx>`:
              &'fcx T
              &'fcx [u8]
              &'fcx str
              *mut FunctionCallInfoBaseData
              AnyArray
              AnyElement
              AnyNumeric
              BOX
            and 36 others
note: required by a bound in `pgrx::callconv::Args::<'a, 'fcx>::next_arg_unchecked`
   --> $HOME/.cargo/registry/src/index.crates.io-6f17d22bba15001f/pgrx-0.12.9/src/callconv.rs:931:41
    |
931 |     pub unsafe fn next_arg_unchecked<T: ArgAbi<'fcx>>(&mut self) -> Option<T> {
    |                                         ^^^^^^^^^^^^ required by this bound in `Args::<'a, 'fcx>::next_arg_unchecked`
    = note: this error originates in the attribute macro `pg_extern` (in Nightly builds, run with -Z macro-backtrace for more info)

Wow, that is a lot of compile errors at the first glance, but if we take a closer look, there are only 3 errors, let's fix them one by one:)

Fix the errors!

the trait bound for<'a> fn(&'a CStr, Oid, i32) -> Vector: FunctionMetadata<_> is not satisfied

Well, the message of this error is quite messy, but there is indeed one useful note:

note: the trait `SqlTranslatable` must be implemented
  --> $HOME/.cargo/registry/src/index.crates.io-6f17d22bba15001f/pgrx-sql-entity-graph-0.12.9/src/metadata/sql_translatable.rs:73:1
   |
73 | pub unsafe trait SqlTranslatable {
   | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
   = help: items from traits can only be used if the trait is implemented and in scope
   = note: the following traits define an item `entity`, perhaps you need to implement one of them:
           candidate #1: `FunctionMetadata`
           candidate #2: `SqlTranslatable`
   = note: this error originates in the attribute macro `pg_extern` (in Nightly builds, run with -Z macro-backtrace for more info)

This note says that we should implement SqlTranslatable for our Vector type, what is trait for? Its doc says that this trait represents "A value which can be represented in SQL", which is kinda confusing. Let's me explain a bit. Since pgrx generates SQL for us, and it would use our Vector type, it requires us to tell it what our type should be called in SQL, a.k.a., translate it to SQL.

In SQL, our Vector type is simply called vector, so here is our implementation:

unsafe impl SqlTranslatable for Vector {
    fn argument_sql() -> Result<SqlMapping, ArgumentError> {
        Ok(SqlMapping::As("vector".into()))
    }

    fn return_sql() -> Result<Returns, ReturnsError> {
        Ok(Returns::One(SqlMapping::As("vector".into())))
    }
}

the trait bound Vector: RetAbi and Vector: ArgAbi are not satisfied

Our Vector type has its memory representation (how it gets laid out in memory), and Postgres has its own way (or rule) to store things in memory. The vector_input() function returns a Vector to Postgres, and the vector_output() function takes a Vector argument, which will be supplied by Postgres:

fn vector_input(input: &CStr, _oid: pg_sys::Oid, type_modifier: i32) -> Vector
fn vector_output(value: Vector) -> CString

Using Vector as function return value and argument requires it to be representable following Postgres's rule. Postgres uses Datum, the binary representation for all the SQL types. So we need to provide a round trip between our Vector type and Datum, through these 2 traits.

Since our Vector type is basically a std::vec::Vec<f64>, and these 2 traits are implemented for Vec, we can simply use these implementations:

impl FromDatum for Vector {
    unsafe fn from_polymorphic_datum(
        datum: pg_sys::Datum,
        is_null: bool,
        typoid: pg_sys::Oid,
    ) -> Option<Self>
    where
        Self: Sized,
    {
        let value = <Vec<f64> as FromDatum>::from_polymorphic_datum(datum, is_null, typoid)?;
        Some(Self { value })
    }
}

impl IntoDatum for Vector {
    fn into_datum(self) -> Option<pg_sys::Datum> {
        self.value.into_datum()
    }

    fn type_oid() -> pg_sys::Oid {
        rust_regtypein::<Self>()
    }
}

unsafe impl<'fcx> ArgAbi<'fcx> for Vector
where
    Self: 'fcx,
{
    unsafe fn unbox_arg_unchecked(arg: ::pgrx::callconv::Arg<'_, 'fcx>) -> Self {
        unsafe { arg.unbox_arg_using_from_datum().expect("expect it to be non-NULL") }
    }
}

unsafe impl BoxRet for Vector {
    unsafe fn box_into<'fcx>(
        self,
        fcinfo: &mut pgrx::callconv::FcInfo<'fcx>,
    ) -> pgrx::datum::Datum<'fcx> {
        match self.into_datum() {
            Some(datum) => unsafe {fcinfo.return_raw_datum(datum)}
            None => fcinfo.return_null()
        }
    }
}

FromDatum and IntoDatum are also implemented because they will be used in the RetAbi and ArgAbi implementations.

Let's see if there are any errors:

$ cargo check
$ echo $?
0

Great, no errors, let's launch it again!

Our second launch!

$ cargo pgrx run
    ...
psql (17.2)
Type "help" for help.

pg_vector_ext=#

Yes, psql starts without any issue! Let's enable our extension and create a table with a vector column:

pg_vector_ext=# CREATE EXTENSION pg_vector_ext;
CREATE EXTENSION
pg_vector_ext=# CREATE TABLE items (v vector(3));
CREATE TABLE

So far so good, now let's insert some data:

pg_vector_ext=# INSERT INTO items values ('[1, 2, 3]');
ERROR:  failed to cast stored dimension [-1] to u16
LINE 1: INSERT INTO items values ('[1, 2, 3]');
                                  ^
DETAIL:
   0: std::backtrace_rs::backtrace::libunwind::trace
             at /rustc/f6e511eec7342f59a25f7c0534f1dbea00d01b14/library/std/src/../../backtrace/src/backtrace/libunwind.rs:116:5
             ...
  46: <unknown>
             at main.c:197:3

Well, Postgres panicked, because the type_modifier argument is -1, which can not be casted into a u16. But how can the type_modifier be -1, wouldn't it be -1 if and only if the type modifier is unknown? The type modifier 3 is clearly stored in the metadata, so it is definitely known.

I was stuck here for a while, and guess what, I am not the only one. IMHO, this is a bug, but not everyone thinks so. Let's accept the truth that it just can be -1, the workaround for this is to create a cast function that casts Vector to Vector, where you can access the stored type modifier:

/// Cast a `vector` to a `vector`, the conversion is meaningless, but we do need
/// to do the dimension check here if we cannot get the `typmod` value in vector
/// type's input function.
#[pgrx::pg_extern(immutable, strict, parallel_safe, requires = ["concrete_type"])]
fn cast_vector_to_vector(vector: Vector, type_modifier: i32, _explicit: bool) -> Vector {
    let expected_dimension = u16::try_from(type_modifier).expect("invalid type_modifier") as usize;
    let dimension = vector.value.len();
    if vector.value.len() != expected_dimension {
        pgrx::error!(
            "mismatched dimension, expected {}, found {}",
            type_modifier,
            dimension
        );
    }

    vector
}

extension_sql!(
    r#"
    CREATE CAST (vector AS vector)
    WITH FUNCTION cast_vector_to_vector(vector, integer, boolean);
    "#,
    name = "cast_vector_to_vector",
    requires = ["concrete_type", cast_vector_to_vector]
);

In our vector_input() function, no dimension check will be performed if the type modifier is unknown:

#[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
fn vector_input(input: &CStr, _oid: pg_sys::Oid, type_modifier: i32) -> Vector {
    let value = match serde_json::from_str::<Vec<f64>>(
        input.to_str().expect("expect input to be UTF-8 encoded"),
    ) {
        Ok(v) => v,
        Err(e) => {
            pgrx::error!("failed to deserialize the input string due to error {}", e)
        }
    };
    let dimension = match u16::try_from(value.len()) {
        Ok(d) => d,
        Err(_) => {
            pgrx::error!("this vector's dimension [{}] is too large", value.len());
        }
    };

    // check the dimension in INPUT function if we know the expected dimension.
    if type_modifier != -1 {
        let expected_dimension = match u16::try_from(type_modifier) {
            Ok(d) => d,
            Err(_) => {
                panic!("failed to cast stored dimension [{}] to u16", type_modifier);
            }
        };

        if dimension != expected_dimension {
            pgrx::error!(
                "mismatched dimension, expected {}, found {}",
                expected_dimension,
                dimension
            );
        }
    }
    // If we don't know the type modifier, do not check the dimension.

    Vector { value }
}

Now everything should really work:

pg_vector_ext=# INSERT INTO items values ('[1, 2, 3]');
INSERT 0 1

INSERT works, let's try SELECT:

pg_vector_ext=# SELECT * FROM items;
       v
---------------
 [1.0,2.0,3.0]
(1 row)

Congratulation! You just added the vector type support for Postgres! Here is the full code of our implementation:

src/vector_type.rs

//! This file defines a `Vector` type.

use pgrx::callconv::ArgAbi;
use pgrx::callconv::BoxRet;
use pgrx::datum::FromDatum;
use pgrx::datum::IntoDatum;
use pgrx::extension_sql;
use pgrx::pg_extern;
use pgrx::pg_sys;
use pgrx::pgrx_sql_entity_graph::metadata::ArgumentError;
use pgrx::pgrx_sql_entity_graph::metadata::Returns;
use pgrx::pgrx_sql_entity_graph::metadata::ReturnsError;
use pgrx::pgrx_sql_entity_graph::metadata::SqlMapping;
use pgrx::pgrx_sql_entity_graph::metadata::SqlTranslatable;
use pgrx::wrappers::rust_regtypein;
use std::ffi::CStr;
use std::ffi::CString;

/// The `vector` type
#[derive(Debug, serde::Deserialize, serde::Serialize)]
#[serde(transparent)]
pub(crate) struct Vector {
    pub(crate) value: Vec<f64>,
}

unsafe impl SqlTranslatable for Vector {
    fn argument_sql() -> Result<SqlMapping, ArgumentError> {
        Ok(SqlMapping::As("vector".into()))
    }

    fn return_sql() -> Result<Returns, ReturnsError> {
        Ok(Returns::One(SqlMapping::As("vector".into())))
    }
}

impl FromDatum for Vector {
    unsafe fn from_polymorphic_datum(
        datum: pg_sys::Datum,
        is_null: bool,
        typoid: pg_sys::Oid,
    ) -> Option<Self>
    where
        Self: Sized,
    {
        let value = <Vec<f64> as FromDatum>::from_polymorphic_datum(datum, is_null, typoid)?;
        Some(Self { value })
    }
}

impl IntoDatum for Vector {
    fn into_datum(self) -> Option<pg_sys::Datum> {
        self.value.into_datum()
    }

    fn type_oid() -> pg_sys::Oid {
        rust_regtypein::<Self>()
    }
}

unsafe impl<'fcx> ArgAbi<'fcx> for Vector
where
    Self: 'fcx,
{
    unsafe fn unbox_arg_unchecked(arg: ::pgrx::callconv::Arg<'_, 'fcx>) -> Self {
        unsafe {
            arg.unbox_arg_using_from_datum()
                .expect("expect it to be non-NULL")
        }
    }
}

unsafe impl BoxRet for Vector {
    unsafe fn box_into<'fcx>(
        self,
        fcinfo: &mut pgrx::callconv::FcInfo<'fcx>,
    ) -> pgrx::datum::Datum<'fcx> {
        match self.into_datum() {
            Some(datum) => unsafe { fcinfo.return_raw_datum(datum) },
            None => fcinfo.return_null(),
        }
    }
}

extension_sql!(
    r#"
CREATE TYPE vector; -- shell type
"#,
    name = "shell_type",
    bootstrap // declare this extension_sql block as the "bootstrap" block so that it happens first in sql generation
);

#[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
fn vector_input(input: &CStr, _oid: pg_sys::Oid, type_modifier: i32) -> Vector {
    let value = match serde_json::from_str::<Vec<f64>>(
        input.to_str().expect("expect input to be UTF-8 encoded"),
    ) {
        Ok(v) => v,
        Err(e) => {
            pgrx::error!("failed to deserialize the input string due to error {}", e)
        }
    };
    let dimension = match u16::try_from(value.len()) {
        Ok(d) => d,
        Err(_) => {
            pgrx::error!("this vector's dimension [{}] is too large", value.len());
        }
    };

    // check the dimension in INPUT function if we know the expected dimension.
    if type_modifier != -1 {
        let expected_dimension = match u16::try_from(type_modifier) {
            Ok(d) => d,
            Err(_) => {
                panic!("failed to cast stored dimension [{}] to u16", type_modifier);
            }
        };

        if dimension != expected_dimension {
            pgrx::error!(
                "mismatched dimension, expected {}, found {}",
                expected_dimension,
                dimension
            );
        }
    }
    // If we don't know the type modifier, do not check the dimension.

    Vector { value }
}

#[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
fn vector_output(value: Vector) -> CString {
    let value_serialized_string = serde_json::to_string(&value).unwrap();
    CString::new(value_serialized_string).expect("there should be no NUL in the middle")
}

#[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
fn vector_modifier_input(list: pgrx::datum::Array<&CStr>) -> i32 {
    if list.len() != 1 {
        pgrx::error!("too many modifiers, expect 1")
    }

    let modifier = list
        .get(0)
        .expect("should be Some as len = 1")
        .expect("type modifier cannot be NULL");
    let Ok(dimension) = modifier
        .to_str()
        .expect("expect type modifiers to be UTF-8 encoded")
        .parse::<u16>()
    else {
        pgrx::error!("invalid dimension value, expect a number in range [1, 65535]")
    };

    dimension as i32
}

#[pg_extern(immutable, strict, parallel_safe, requires = [ "shell_type" ])]
fn vector_modifier_output(type_modifier: i32) -> CString {
    CString::new(format!("({})", type_modifier)).expect("no NUL in the middle")
}

// create the actual type, specifying the input and output functions
extension_sql!(
    r#"
CREATE TYPE vector (
    INPUT = vector_input,
    OUTPUT = vector_output,
    TYPMOD_IN = vector_modifier_input,
    TYPMOD_OUT = vector_modifier_output,
    STORAGE = external 
);
"#,
    name = "concrete_type",
    creates = [Type(Vector)],
    requires = [
        "shell_type",
        vector_input,
        vector_output,
        vector_modifier_input,
        vector_modifier_output
    ]
);

/// Cast a `vector` to a `vector`, the conversion is meaningless, but we do need
/// to do the dimension check here if we cannot get the `typmod` value in vector
/// type's input function.
#[pgrx::pg_extern(immutable, strict, parallel_safe, requires = ["concrete_type"])]
fn cast_vector_to_vector(vector: Vector, type_modifier: i32, _explicit: bool) -> Vector {
    let expected_dimension = u16::try_from(type_modifier).expect("invalid type_modifier") as usize;
    let dimension = vector.value.len();
    if vector.value.len() != expected_dimension {
        pgrx::error!(
            "mismatched dimension, expected {}, found {}",
            type_modifier,
            dimension
        );
    }

    vector
}

extension_sql!(
    r#"
    CREATE CAST (vector AS vector)
    WITH FUNCTION cast_vector_to_vector(vector, integer, boolean);
    "#,
    name = "cast_vector_to_vector",
    requires = ["concrete_type", cast_vector_to_vector]
);