stencilpy

Extract structured data from Excel files using YAML schema definitions into dynamically-generated Pydantic models.

Installation

pip install stencilpy

The stencil open command serves the bundled editor UI locally and opens it in your browser.

Quick Start

from stencilpy import Stencil

# Load a schema
lab = Stencil("lab_report.stencil.yaml")

# Extract data — version auto-detected via discriminator
report = lab.extract("january_lab.xlsx")
print(report.patient_name)
print(report.model_dump())

Schema Format

Create a .stencil.yaml file:

name: lab_report
description: Monthly lab report

discriminator:
  cells:
    - A1

versions:
  "v2.0":
    fields:
      patient_name:
        cell: B3
      sample_date:
        cell: B4
        type: datetime
      readings:
        range: D5:D
        type: list[float]
      report_version:
        cell: header:right
      footer_note:
        cell: footer:center

Scalar cell references can also target worksheet headers and footers:

• header:left
• header:center
• header:right
• footer:left
• footer:center
• footer:right
• Sheet1!header:first:right
• Sheet1!footer:even:center

These references also work in discriminator.cells, which is useful when a workbook version is printed in the page header/footer instead of a normal cell.

Header And Footer References

Use header/footer refs anywhere a scalar cell ref is accepted.

Example: extract version text and report metadata from the page chrome.

name: lab_report
description: Monthly lab report

discriminator:
  cells:
    - A1

versions:
  "v2.0":
    fields:
      patient_name:
        cell: B3
      report_version:
        cell: header:right
      report_title:
        cell: header:center
      generated_by:
        cell: footer:left
      footer_note:
        cell: footer:center

If the workbook uses separate first-page or even-page headers/footers, include the page selector:

versions:
  "v2.0":
    fields:
      first_page_title:
        cell: header:first:center
      even_page_version:
        cell: footer:even:right
      cover_sheet_version:
        cell: Cover!header:first:right

Supported formats:

• header:left
• header:center
• header:right
• header:first:left
• header:even:center
• footer:right
• footer:first:center
• Sheet1!header:right
• Sheet1!footer:even:left

Header-Based Version Detection

If a workbook stores its version in a header or footer instead of a normal cell, add those refs to discriminator.cells.

name: lab_report
description: Monthly lab report

discriminator:
  cells:
    - A1
    - header:right
    - Cover!footer:first:center

versions:
  "v1.0":
    fields:
      patient_name:
        cell: A3
  "v2.0":
    fields:
      patient_name:
        cell: B3

stencilpy will check each discriminator ref in order until one matches a known version key.

Phlo Export

stencil phlo converts a schema into the dlt ingestion asset, Pandera schema and dbt models a Phlo project needs to land and model the same workbooks:

stencil phlo lab_report.stencil.yaml --out ./my-phlo-project

Pass a directory instead of a file to generate artifacts for every .stencil.yaml in it:

stencil phlo ./schemas --out ./my-phlo-project

Run it against a Phlo project (or an empty directory you plan to use as one) and it writes:

dbt tests and column docs sit next to each model as .yml files, and every generated file belongs to a single schema, so several schemas can share one project without overwriting each other. Schemas with the same name are rejected.

The generated asset reads workbooks from data/<table> by default (override with STENCIL_INPUT_DIR or --input-dir) and lands one raw row per workbook, keyed by record_id (<partition date>:<relative path>), so re-running a partition is idempotent. Workbooks whose layout matches no schema version fail the run with a VersionError. --table, --domain and --input-dir only apply when generating a single schema file.

Scalar fields keep their stencil types. list, dict and table fields land as JSON text because Phlo's dlt integration normalises nested values into child tables, which the raw Iceberg table cannot represent; the generated silver models explode them back into one row per entry. Computed fields have no declared type, so they land as text.

Existing files are not overwritten unless --force is passed. --table and --domain override the generated table and workflow names:

stencil phlo lab_report.stencil.yaml --out . --table lab_reports --domain labs --force

The same conversion is available from Python:

from stencilpy.phlo import write_phlo_files

write_phlo_files("lab_report.stencil.yaml", "./my-phlo-project")

Dialects

The dbt models target Trino by default, the engine Phlo's dbt profile uses. Pass --dialect to generate them for another engine:

stencil phlo lab_report.stencil.yaml --out . --dialect duckdb

Only the dbt SQL depends on the dialect: the dlt asset, Pandera schema, sources.yml and schema.yml are the same for every engine. build_phlo_files/write_phlo_files also accept a PhloDialect instance, so another engine can be supported by subclassing PhloDialect (column types, identifier quoting, JSON extraction and the join that explodes a collection).

Note that Trino casts JSON array elements individually, turning uncastable elements into NULL, while DuckDB raises a conversion error instead.