How AkaXtract Learns Your Documents
AkaXtract does not train or fine-tune a model on your documents. You do not prepare a training data set, and you do not wait for a training run. You describe what to extract in a Pipeline, and a change takes effect when you publish a new version.
What You Configure
| Setting | What it controls |
|---|---|
| AI model | The model that processes your documents. Accuracy, language coverage, and page limits can differ between models |
| Classification prompt and schema | How each page is assigned to a document type, and how pages are grouped into documents |
| Extraction prompt (per document type) | Which fields to find, how to treat missing values, and how to read tables |
| Extraction JSON schema (per document type) | Field names, data types, and the structure of the result, such as header fields and line items |
| Post-process script (optional) | Corrections to values after extraction |
| Approval rules | Which results finish automatically and which go to a reviewer |
Improve Accuracy
- Create the Pipeline from a sample that represents your real documents.
- Check that the generated fields, names, and types match what your receiving system needs.
- Test with sample documents and compare the data with the source.
- If a field is wrong or missing, change the extraction prompt or schema of that document type. Describe the required fields, how to handle missing values, and table data more precisely. Use a post-process script for simple fixes. If results are still weak, select a different AI model and test again.
- Test again, then publish.
- Keep
Auto-Approve Resultsoff while you validate. Turn it on, with a threshold, when you trust the results. See Approval Rules.
What Review Corrections Do
A correction in Review fixes the result of that one document. It does not train the model and does not change the Pipeline. To apply a fix to future documents, change the prompt or schema in a new Draft, then test and publish it.