apache_beam.ml.rag.ingestion.postgres_common module

apache_beam.ml.rag.ingestion.postgres_common.chunk_embedding_fn(chunk: Chunk) → str

Convert embedding to PostgreSQL array string.

Formats dense embedding as a PostgreSQL-compatible array string. Example: [1.0, 2.0] -> ‘{1.0,2.0}’

Parameters:

chunk – Input Chunk object.

Returns:

PostgreSQL array string representation of the embedding.

Return type:

str

Raises:

ValueError – If chunk has no dense embedding.

class apache_beam.ml.rag.ingestion.postgres_common.ColumnSpec(column_name: str, python_type: Type, value_fn: Callable[[Chunk], Any], sql_typecast: str | None = None)

Bases: object

Specification for mapping Chunk fields to SQL columns for insertion.

Defines how to extract and format values from Chunks into database columns, handling the full pipeline from Python value to SQL insertion.

The insertion process works as follows: - value_fn extracts a value from the Chunk and formats it as needed - The value is stored in a NamedTuple field with the specified python_type - During SQL insertion, the value is bound to a ? placeholder

column_name

The column name in the database table.

Type:

str

python_type

Python type for the NamedTuple field that will hold the value. Must be compatible with must be compatible with RowCoder.

Type:

Type

value_fn

Function to extract and format the value from a Chunk. Takes a Chunk and returns a value of python_type.

Type:

Callable[[apache_beam.ml.rag.types.Chunk], Any]

sql_typecast

Optional SQL type cast to append to the ? placeholder. Common examples: - “::float[]” for vector arrays - “::jsonb” for JSON data

Type:

str | None

Examples

Basic text column (uses standard JDBC type mapping): >>> ColumnSpec.text( … column_name=”content”, … value_fn=lambda chunk: chunk.content.text … ) # Results in: INSERT INTO table (content) VALUES (?)

Vector column with explicit array casting: >>> ColumnSpec.vector( … column_name=”embedding”, … value_fn=lambda chunk: ‘{’ + … ‘,’.join(map(str, chunk.embedding.dense_embedding)) + ‘}’ … ) # Results in: INSERT INTO table (embedding) VALUES (?::float[]) # The value_fn formats [1.0, 2.0] as ‘{1.0,2.0}’ for PostgreSQL array

Timestamp from metadata with explicit casting: >>> ColumnSpec( … column_name=”created_at”, … python_type=str, … value_fn=lambda chunk: chunk.metadata.get(“timestamp”), … sql_typecast=”::timestamp” … ) # Results in: INSERT INTO table (created_at) VALUES (?::timestamp) # Allows inserting string timestamps with proper PostgreSQL casting

Factory Methods:

text: Creates a text column specification (no type cast). integer: Creates an integer column specification (no type cast). float: Creates a float column specification (no type cast). vector: Creates a vector column specification with float[] casting. jsonb: Creates a JSONB column specification with jsonb casting.

column_name: str

python_type: Type

value_fn: Callable[[Chunk], Any]

sql_typecast: str | None = None

property placeholder: str

Get SQL placeholder with optional typecast.

classmethod text(column_name: str, value_fn: Callable[[Chunk], Any]) → ColumnSpec

Create a text column specification.

classmethod integer(column_name: str, value_fn: Callable[[Chunk], Any]) → ColumnSpec

Create an integer column specification.

classmethod float(column_name: str, value_fn: Callable[[Chunk], Any]) → ColumnSpec

Create a float column specification.

classmethod vector(column_name: str, value_fn: ~typing.Callable[[~apache_beam.ml.rag.types.Chunk], ~typing.Any] = <function chunk_embedding_fn>) → ColumnSpec

Create a vector column specification.

classmethod jsonb(column_name: str, value_fn: Callable[[Chunk], Any]) → ColumnSpec

Create a JSONB column specification.

class apache_beam.ml.rag.ingestion.postgres_common.ColumnSpecsBuilder

Bases: object

Builder for ColumnSpec’s with chainable methods.

static with_defaults() → ColumnSpecsBuilder

Add all default column specifications.

with_id_spec(column_name: str = 'id', python_type: ~typing.Type = <class 'str'>, convert_fn: ~typing.Callable[[str], ~typing.Any] | None = None, sql_typecast: str | None = None) → ColumnSpecsBuilder

Add ID ColumnSpec with optional type and conversion.

Parameters:

• column_name – Name for the ID column (defaults to “id”)

• python_type – Python type for the column (defaults to str)

• convert_fn – Optional function to convert the chunk ID If None, uses ID as-is

• sql_typecast – Optional SQL type cast

Returns:

Self for method chaining

Example

>>> builder.with_id_spec(
...     column_name="doc_id",
...     python_type=int,
...     convert_fn=lambda id: int(id.split('_')[1])
... )

with_content_spec(column_name: str = 'content', python_type: ~typing.Type = <class 'str'>, convert_fn: ~typing.Callable[[str], ~typing.Any] | None = None, sql_typecast: str | None = None) → ColumnSpecsBuilder

Add content ColumnSpec with optional type and conversion.

Parameters:

• column_name – Name for the content column (defaults to “content”)

• python_type – Python type for the column (defaults to str)

• convert_fn – Optional function to convert the content text If None, uses content text as-is

• sql_typecast – Optional SQL type cast

Returns:

Self for method chaining

Example

>>> builder.with_content_spec(
...     column_name="content_length",
...     python_type=int,
...     convert_fn=len  # Store content length instead of content
... )

with_metadata_spec(column_name: str = 'metadata', python_type: ~typing.Type = <class 'str'>, convert_fn: ~typing.Callable[[~typing.Dict[str, ~typing.Any]], ~typing.Any] | None = None, sql_typecast: str | None = '::jsonb') → ColumnSpecsBuilder

Add metadata ColumnSpec with optional type and conversion.

Parameters:

• column_name – Name for the metadata column (defaults to “metadata”)

• python_type – Python type for the column (defaults to str)

• convert_fn – Optional function to convert the metadata dictionary If None and python_type is str, converts to JSON string

• sql_typecast – Optional SQL type cast (defaults to “::jsonb”)

Returns:

Self for method chaining

Example

>>> builder.with_metadata_spec(
...     column_name="meta_tags",
...     python_type=list,
...     convert_fn=lambda meta: list(meta.keys()),
...     sql_typecast="::text[]"
... )

with_embedding_spec(column_name: str = 'embedding', convert_fn: Callable[[List[float]], Any] | None = None) → ColumnSpecsBuilder

Add embedding ColumnSpec with optional conversion.

Parameters:

• column_name – Name for the embedding column (defaults to “embedding”)

• convert_fn – Optional function to convert the dense embedding values If None, uses default PostgreSQL array format

Returns:

Self for method chaining

Example

>>> builder.with_embedding_spec(
...     column_name="embedding_vector",
...     convert_fn=lambda values: '{' + ','.join(f"{x:.4f}"
...       for x in values) + '}'
... )

add_metadata_field(field: str, python_type: Type, column_name: str | None = None, convert_fn: Callable[[Any], Any] | None = None, default: Any | None = None, sql_typecast: str | None = None) → ColumnSpecsBuilder

“”Add a ColumnSpec that extracts and converts a field from chunk metadata.

Parameters:

• field – Key to extract from chunk metadata

• python_type – Python type for the column (e.g. str, int, float)

• column_name – Name for the column (defaults to metadata field name)

• convert_fn – Optional function to convert the extracted value to desired type. If None, value is used as-is

• default – Default value if field is missing from metadata

• sql_typecast – Optional SQL type cast (e.g. “::timestamp”)

Returns:

Self for chaining

Examples

Simple string field: >>> builder.add_metadata_field(“source”, str)

Integer with default:

>>> builder.add_metadata_field(
...     field="count",
...     python_type=int,
...     column_name="item_count",
...     default=0
... )

Float with conversion and default:

>>> builder.add_metadata_field(
...     field="confidence",
...     python_type=intfloat,
...     convert_fn=lambda x: round(float(x), 2),
...     default=0.0
... )

Timestamp with conversion and type cast:

>>> builder.add_metadata_field(
...     field="created_at",
...     python_type=intstr,
...     convert_fn=lambda ts: ts.replace('T', ' '),
...     sql_typecast="::timestamp"
... )

add_custom_column_spec(spec: ColumnSpec) → ColumnSpecsBuilder

Add a custom ColumnSpec to the builder.

Use this method when you need complete control over the ColumnSpec , including custom value extraction and type handling.

Parameters:

spec – A ColumnSpec instance defining the column name, type, value extraction, and optional SQL type casting.

Returns:

Self for method chaining

Examples

Custom text column from chunk metadata:

>>> builder.add_custom_column_spec(
...     ColumnSpec.text(
...         name="source_and_id",
...         value_fn=lambda chunk:         ...             f"{chunk.metadata.get('source')}_{chunk.id}"
...     )
... )

build() → List[ColumnSpec]

Build the final list of column specifications.

class apache_beam.ml.rag.ingestion.postgres_common.ConflictResolution(on_conflict_fields: str | List[str], action: Literal['UPDATE', 'IGNORE'] = 'UPDATE', update_fields: List[str] | None = None)

Bases: object

Specification for how to handle conflicts during insert.

Configures conflict handling behavior when inserting records that may violate unique constraints.

on_conflict_fields

Field(s) that determine uniqueness. Can be a single field name or list of field names for composite constraints.

Type:

str | List[str]

action

How to handle conflicts - either “UPDATE” or “IGNORE”. UPDATE: Updates existing record with new values. IGNORE: Skips conflicting records.

Type:

Literal[‘UPDATE’, ‘IGNORE’]

update_fields

Optional list of fields to update on conflict. If None, all non-conflict fields are updated.

Type:

List[str] | None

Examples

Simple primary key:

>>> ConflictResolution("id")

Composite key with specific update fields:

>>> ConflictResolution(
...     on_conflict_fields=["source", "timestamp"],
...     action="UPDATE",
...     update_fields=["embedding", "content"]
... )

Ignore conflicts:

>>> ConflictResolution(
...     on_conflict_fields="id",
...     action="IGNORE"
... )

on_conflict_fields: str | List[str]

action: Literal['UPDATE', 'IGNORE'] = 'UPDATE'

update_fields: List[str] | None = None

maybe_set_default_update_fields(columns: List[str])

get_conflict_clause() → str

Get conflict clause with update fields.