Skip to content

Exporting Results

The export tool writes any pipeline output to disk in a choice of formats. It works as the final task in any pipeline and is the standard way to produce files from an extraction or LLM workflow.

Tools used: export
Requires: Nothing — the examples below use mock data so no LLM key is needed

Supported formats

format value Output Best for
json Pretty-printed JSON file Machine consumption, debugging
markdown Markdown table Human review, audit trail
csv Comma-separated values Spreadsheet import
xlsx Excel workbook Finance team handoff

Basic export — JSON

pipelines/export_json.yaml
pipeline:
  id: export_json_example
  goal: "Extract data and write it to JSON"

  tasks:
    - id: produce_data
      tool: mock
      inputs:
        revenue: 391035
        gross_profit: 180683
        net_income: 93736
        eps_diluted: 6.08

    - id: write_json
      tool: export
      inputs:
        data: "{{produce_data.output.inputs}}"
        format: json
        filename: extraction_result
        output_dir: outputs

import asyncio
from trellis.models.pipeline import Pipeline
from trellis.execution.orchestrator import Orchestrator

async def main():
    pipeline = Pipeline.from_yaml_file("pipelines/export_json.yaml")
    orch = Orchestrator()
    result = await orch.run_pipeline(pipeline)

    export = result.outputs["write_json"]
    print(f"Written to: {export['path']}")   # outputs/extraction_result.json
    print(f"Size: {export['size']} bytes")

asyncio.run(main())

Output file (outputs/extraction_result.json):

{
  "revenue": 391035,
  "gross_profit": 180683,
  "net_income": 93736,
  "eps_diluted": 6.08
}

Task output:

{
    "status": "success",
    "format": "json",
    "filename": "extraction_result.json",
    "path": "/absolute/path/outputs/extraction_result.json",
    "size": 89
}

Markdown export with metadata

Markdown export is useful for financial spreads — it renders the fields as a labelled table and prepends document metadata as a header block:

pipelines/export_markdown.yaml
pipeline:
  id: export_markdown_example
  goal: "Export extraction result as a Markdown spread"

  params:
    ticker:
      type: string
      default: AAPL
    period_end:
      type: string
      default: "2024-09-30"

  tasks:
    - id: produce_data
      tool: mock
      inputs:
        revenue: 391035
        gross_profit: 180683
        operating_income: 123216
        net_income: 93736
        eps_diluted: 6.08

    - id: write_markdown
      tool: export
      inputs:
        data: "{{produce_data.output.inputs}}"
        format: markdown
        filename: "{{params.ticker}}_income_statement"
        output_dir: outputs
        metadata:
          company: Apple Inc.
          ticker: "{{params.ticker}}"
          period_end: "{{params.period_end}}"
          currency: USD
          units: millions
          audited: true
        periods:
          - label: "FY2024"
            period_end: "{{params.period_end}}"

Output file (outputs/AAPL_income_statement.md):

| Field | Value |
| --- | --- |
| Company | Apple Inc. |
| Ticker | AAPL |
| Period End | 2024-09-30 |
| Currency | USD |
| Units | millions |
| Audited | true |

| Field | FY2024 |
| --- | --- |
| revenue | 391,035 |
| gross_profit | 180,683 |
| operating_income | 123,216 |
| net_income | 93,736 |
| eps_diluted | 6.08 |

CSV and XLSX export

CSV and XLSX exports write a header row followed by one data row:

- id: write_csv
  tool: export
  inputs:
    data: "{{extract.output}}"
    format: csv
    filename: extraction
    output_dir: outputs

- id: write_xlsx
  tool: export
  inputs:
    data: "{{extract.output}}"
    format: xlsx
    filename: extraction
    output_dir: outputs

Exporting multiple formats from one pipeline

Run export tasks in parallel since they depend on the same upstream data and have no dependency on each other:

pipelines/multi_format_export.yaml
pipeline:
  id: multi_format_export
  goal: "Export extraction result to JSON, Markdown, and CSV simultaneously"

  tasks:
    - id: extract
      tool: mock
      inputs:
        revenue: 391035
        net_income: 93736
        eps_diluted: 6.08

    - id: export_json
      tool: export
      inputs:
        data: "{{extract.output.inputs}}"
        format: json
        filename: result
        output_dir: outputs

    - id: export_md
      tool: export
      inputs:
        data: "{{extract.output.inputs}}"
        format: markdown
        filename: result
        output_dir: outputs

    - id: export_csv
      tool: export
      inputs:
        data: "{{extract.output.inputs}}"
        format: csv
        filename: result
        output_dir: outputs

export_json, export_md, and export_csv all depend on extract and have no dependency on each other, so they run concurrently in wave 2.

import asyncio
from trellis.models.pipeline import Pipeline
from trellis.execution.orchestrator import Orchestrator

async def main():
    pipeline = Pipeline.from_yaml_file("pipelines/multi_format_export.yaml")
    orch = Orchestrator()
    result = await orch.run_pipeline(pipeline)

    for task_id in ("export_json", "export_md", "export_csv"):
        out = result.outputs[task_id]
        print(f"{out['format']:8}  →  {out['path']}")

asyncio.run(main())

json      →  /path/to/outputs/result.json
markdown  →  /path/to/outputs/result.md
csv       →  /path/to/outputs/result.csv

result.waves_executed is 2: wave 1 runs extract, wave 2 runs the three exports concurrently.

Chaining from extraction

In a real pipeline, replace the mock task with an extract_fields task. The output shape is the same — a flat dict of field name → value:

- id: extract
  tool: extract_fields
  inputs:
    document: "{{select_pages.output}}"
    schema: "{{schema.output}}"
    period_end: "{{params.period_end}}"
    section_filter: face

- id: export_json
  tool: export
  inputs:
    data: "{{extract.output}}"    # same: flat dict of field values
    format: json
    filename: "{{params.ticker}}_extraction"
    output_dir: outputs

See SEC Filing Extraction for the complete pipeline that includes the fetch, ingest, and select steps before this.

Next steps