Skip to content

Quickstart

Install docsynth and generate your first synthetic documents — with a computed golden dataset — using the studio. This walk-through uses the local target: no cloud account, no API keys.

Prerequisites

  • Python 3.12+ and git.
  • Chromium for PDF rendering. docsynth renders PDFs through headless Chromium (via Playwright); one command installs it below. Skip it only if you will generate --format html exclusively.
  • (Optional, cloud) the gcloud CLI and a GCP project — only if you later use the gcp target. Not needed here.

Install docsynth

# 1. Create and activate a virtual environment
python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate

# 2. Install docsynth with the studio wizard extra
pip install 'docsynth[studio]'

# 3. Install the headless Chromium the PDF renderer uses
playwright install chromium

git clone https://github.com/kashodev/docsynth.git
cd docsynth
python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install -e '.[studio]'
playwright install chromium

Either way you get the docsynth command:

docsynth --help
docsynth studio --help

Extras

  • '.[studio]' adds arrow-key prompts (questionary). Plain pip install -e . still works — the wizard falls back to numbered menus.
  • '.[gcp]' / '.[aws]' add the cloud backends (gs:// / firestore:// / bigquery:// / s3://); '.[anthropic]' adds the Claude provider; '.[all]' is everything. None are needed for this local quickstart.

Run the studio

docsynth studio is an interactive wizard. Launch it and it walks you through the flow — target → project → pack → step → confirm — prompting only for what you haven't already passed as a flag:

docsynth studio

Choose:

  1. Target → local. Runs on this machine, synchronously; no keys.
  2. Project → type a new name. The studio creates a local workspace (a blobs/ dir for documents + shards and a runs.db for run state) and saves it to ~/.docsynth/projects.yaml.
  3. Pack → invoice (the sole installed pack, so it's auto-selected).
  4. Steps — the studio runs the three in order:
    • catalog — build a product catalogue. Keep it small to start (e.g. 5 companies × 10 products); the default mix is procedural, so no LLM and no keys.
    • pdfs — how many documents to render (e.g. 20), and pdf or html.
    • export — write the golden dataset (every field value, exact to the cent) to DuckDB / Parquet.
  5. Confirm — it runs locally and prints where everything landed.

Prefer one command? Every prompt has a flag, so the same run is scriptable — see Studio → Steps for the fully-flagged (CI) form.

What you get

  • Rendered documents in the workspace blobs/ (the PDFs/HTML).
  • A golden dataset from the export step — the ground truth for every document, which is the point: score an OCR/extraction model against it. See Data model & the golden set.

Next steps