Upload Ground Truth Documents

Want to know whether a model actually got an extraction right? You need a "right answer" to compare against. The first half of that — the source document — lives here.

A Ground Truth Document is a PDF you've already parsed yourself (or had a human parse), tagged with the Document Type it belongs to. Once it's uploaded, Librari Evals OCRs the text, stores the file, and lines it up to be matched with one or more Ground Truth Responses. From there it can land in Test Sets and be benchmarked against any model.

What you'll need

• One or more PDF files. Other formats may work, but PDF is the supported path.
• The Document Type you want to associate them with — create it from Schema Overview first if you haven't.
• A few seconds per document for OCR to run on save.

Walkthrough

1. 1. Open Ground Truth Documents

   From the admin sidebar, navigate to /admin/collections/ground-truth-documents. The header reads Ground Truth Documents with a brief description: Known-good parsed documents for evaluation.

   

2. 2. Pick single or bulk

   For one document, click Create New. For many, click Bulk Upload — that's the right call when you're seeding a fresh corpus or backfilling.

3. 3. Pick the file and Document Type

   In the create form, select the PDF and the Document Type it belongs to. Document Type matters — it's how schemas know which documents are eligible for scoring against them.

4. 4. Save — OCR runs automatically

   Hit Save. Behind the scenes the ocrBeforeChange hook runs the file through the OCR service and stores the extracted text alongside the upload. Save completes once OCR is done.

   Large or scanned-image PDFs take longer. If save seems slow, that's the OCR.

After upload

Each Ground Truth Document is a file with text. To make it useful for scoring you need to attach at least one Ground Truth Response — the human-verified expected output for a given Request Schema Type.

After that, group documents into Test Sets and run them via Run a Manual Test or Schedule Recurring Tests.

Common errors

• OCR failures — if the PDF is image-only and OCR can't extract text, the save will return an error. You'll need a re-scan or a different copy of the file.
• Wrong Document Type — schemas are scoped to a Document Type. If you tag a contract as Invoice, none of your contract schemas will see it.

Related

• Define Expected Responses for Ground Truth — the natural next step.
• Assemble a Test Set — once you have a few documents and responses.