@lf/document-ai

The ocr slot on Google Document AI: the text of a scanned PDF or an image. With it, extract(…) reads a deck that is only pictures of pages, and a plugin's ctx.ocr reads the PDFs it fetches.

It is the platform's own Document AI reader (packages/ocr-google), served as a module:

• a file of up to 15 pages and 20 MB goes to the processor's online endpoint;
• a longer one is uploaded to your Cloud Storage bucket, processed in batch, read back and deleted.

lf asks it only of a file that needs it: an image, or a PDF with no usable text layer of its own (lf reads a born-digital PDF's text itself, by the platform's rule).

Install

1. Install @listen-fire/document-ai into the project's externals folder and keep lf-external.yaml (shipped in the package) beside it, as externals/document_ai/lf-external.yaml.

2. Name it in lf.yaml:

   core:
     ocr: document_ai

3. Run lf connect document_ai to give it its settings. They go to the project's secrets file (or LF_SECRET__DOCUMENT_AI__<SETTING> in containers and CI), never into lf.yaml.

Settings

| setting | meaning | |---|---| | project | the Google Cloud project the OCR processor is in | | location | where the processor lives: eu or us | | processor_id | the processor's id (Document AI → Processors; an Enterprise Document OCR processor) | | bucket | a Cloud Storage bucket for files past 15 pages or 20 MB. Empty: such a file is refused, naming the setting | | client_email, private_key | a service account (the key is secret) with Document AI and the bucket's object roles. Leave both empty to use the environment's own Google credentials (a workload identity, gcloud auth application-default login) | | price_per_page | what a page costs you, in dollars. Default 0.0015, the list price ($1.50 per 1,000 pages) | | base_url, storage_url | other endpoints speaking Document AI's REST API and Cloud Storage's (a proxy); empty for Google's |

Cost

Every read reports its cost: the pages Document AI read times price_per_page, labelled document-ai/ocr. lf adds it to the run's spend, so maxRunCostUsd counts it. The module publishes the price (ocr.prices), so lf refuses before reading a file whose pages alone cost more than the run has left.

Tests

pnpm test in this folder starts the module as the process lf starts and runs the ocr slot's conformance kit and the module's own cases against a fake Document AI and Cloud Storage, driven by Google's own clients. A generated service account key signs its own tokens, so nothing reaches Google.