- Overview
- Model building
- Model evaluation
- Model deployment
- Consuming models
- Developer
- Frequently asked questions
Design-time
Frequently asked questions about design-time in IXP in Flow, covering Extract node limits, fields, annotations, and model performance.
Are there any design-time limits for the Extract node?
Design-time limits include:
- Number of fields per field group: up to 32 fields.
- Number of characters per field name, field group name, or field type name: up to 64 characters.
- Number of characters per instruction: up to 4096 characters. This limit applies to all instructions, including the taxonomy-level overall extraction instructions, field group instructions, field instructions, and field type instructions.
- Total size of documents: up to 200 MB. The limit applies to the total size of the documents, not to the number of documents. This limit helps maintain optimal design-time performance and control design-time costs, as design-time processing is not billable.
Can I create generic fields in the Extract node?
No, there is no concept of generic fields in the Extract node. All fields need to belong to a field group. Each field group represents a separate table, where fields mean table columns. However, it is possible to define a parent-level field group representing a document type, and add more general fields as a table under this field group.
Does the model of an Extract node improve performance with labeling or annotations?
No, the model does not yet learn directly from user annotations. The way to improve the performance of your model is to modify your prompt instructions or model configuration. A fine-tunable model that learns from annotations will be available in a future release.
Why do I need to annotate documents if it does not improve the performance of the model?
It is still important to validate and annotate documents to establish ground truth for the performance evaluation of the model, which is captured in the built-in performance metrics in the Eval runs tab.
How many documents do I need to upload to test my prompt instructions?
You should upload a representative sample of your documents to cover all possible variations. This should be at least 20-30 different documents. You are recommended to review the initial predictions on 2 to 3 documents and decide whether you need to modify your prompt instructions before gradually increasing the test sample to 10, 20, 30, X documents, where this will be an iterative process.
How many documents do I need to annotate to get reliable performance metrics?
You are recommended to validate at least 10 documents or 10 examples, or both, for each field, so that the model can provide you with meaningful validation statistics. This is the minimum number of annotations and will depend on the specific use case requirements.
- Are there any design-time limits for the Extract node?
- Can I create generic fields in the Extract node?
- Does the model of an Extract node improve performance with labeling or annotations?
- Why do I need to annotate documents if it does not improve the performance of the model?
- How many documents do I need to upload to test my prompt instructions?
- How many documents do I need to annotate to get reliable performance metrics?