Core Modules

Table of Content

ReadAI

Intelligent document processing for extracting structured data from files.

ReadAI is Amantra’s Intelligent Document Processing (IDP) engine. It uses machine learning to automatically extract structured data from any type of document — eliminating manual data entry.

What Problems Does ReadAI Solve?

Organizations receive thousands of documents every day — invoices, purchase orders, contracts, ID documents, forms, and more. Extracting data from these manually is:

Slow — Each document can take minutes of human time
Error-prone — Human data entry has a high error rate
Expensive — Manual processing at scale requires large teams

ReadAI automates this completely. Documents are processed in seconds with high accuracy.

Supported Document Types

Financial — Invoices, purchase orders, receipts, bank statements
Identity — Passports, national IDs, driving licenses
Legal — Contracts, agreements, NDAs
Healthcare — Medical forms, prescriptions, lab reports
HR — Resumes, offer letters, employee forms
Custom — Any structured or semi-structured document

How ReadAI Works

Upload Document → AI Classification → Data Extraction → Validation → Output

Upload — Documents arrive via email, API, file upload, or SharePoint/Google Drive.
Classification — ReadAI identifies what type of document it is.
Extraction — The trained model extracts the relevant fields.
Validation — Rules check extracted data for completeness and accuracy.
Output — Structured data is sent downstream (to a workflow, database, or API).

Key Features

Pre-trained Models — Ready-to-use models for common document types
Custom Model Training — Train on your own document templates
Multi-format Support — PDF, JPEG, PNG, TIFF, Word, Excel
Multi-language — Extract data from documents in multiple languages
Confidence Scores — Each extracted field includes a confidence percentage
Human Review Queue — Low-confidence extractions are flagged for human review
Audit Trail — Every extraction is logged with the original document

In This Section

Document Extraction — How to process documents
Models — Manage extraction models
Predictions — Review and validate extracted data

Document Extraction

Document extraction is the core operation of ReadAI — automatically reading a document and pulling out the data you need.

Uploading a Document

You can send documents to ReadAI in three ways:

#### 1. Manual Upload (UI)

Go to ReadAI in the left navigation.
Click + Upload Document.
Drag and drop or browse to select your file.
Choose the extraction model to use.
Click Process.

#### 2. Via Workflow (Synapse)

Use the Read AI activity inside a Synapse workflow to automatically process documents received via email, webhook, or file system.

#### 3. Via API

Send documents programmatically using the ReadAI REST API. See API Reference → Document Intelligence.

Choosing an Extraction Model

Select the model that matches your document type:

Invoice Model — Vendor invoices, billing documents
PO Model — Purchase orders
ID Document Model — Passports, national IDs
Custom Model — Documents specific to your business

Don’t have a model for your document type? See Models to train a custom model.

Extraction Output

After processing, ReadAI returns a structured JSON object with:

Confidence Scores

Each field extraction includes a confidence score (0.0 to 1.0):

0.90+ — High confidence — reliable extraction
0.70–0.89 — Medium confidence — review recommended
Below 0.70 — Low confidence — flagged for human review

Extraction Models

An Extraction Model is a trained machine learning model that knows how to read a specific document type and extract the fields you care about.

Pre-Built Models

Amantra ships with pre-trained models for common document types. These are ready to use immediately with no configuration.

Invoice — Vendor, invoice number, date, line items, totals, tax, payment terms
Purchase Order — PO number, buyer, vendor, items, quantities, prices
Identity Document — Full name, date of birth, ID number, expiry date, nationality
Receipt — Merchant, date, items, total, payment method
Bank Statement — Account number, bank, transactions, opening/closing balance

Training a Custom Model

If your documents have a unique layout, you can train a custom extraction model.

#### Step 1: Define Fields

Go to ReadAI → Models → + New Model.
Enter a model name and description.
Define the fields you want to extract (name, type, required/optional).

#### Step 2: Upload Training Documents

Upload at least 20–50 sample documents for training.
More samples = higher accuracy.
Documents should represent the range of variations you expect (different fonts, layouts, handwriting if applicable).

#### Step 3: Annotate

For each training document, click on the area of the document that contains each field. Amantra will learn to find those fields based on position, context, and surrounding text.

#### Step 4: Train

Click Start Training.
Training typically takes 5–30 minutes depending on the number of documents.
You’ll receive a notification when training is complete.

#### Step 5: Test and Validate

Upload test documents that were not in the training set.
Review extraction accuracy and make corrections.
Optionally run another training round with corrected documents.

#### Step 6: Publish

When satisfied with accuracy, click Publish. The model is now available for use in workflows and the extraction UI.

Model Versioning

Every time you retrain a model, a new version is created. You can:

Roll back to a previous version if a new version performs worse.
A/B test two versions on real documents.
Retire old versions when no longer needed.

Model Performance Metrics

Accuracy — % of fields correctly extracted across test documents
Precision — How often an extracted value is correct
Recall — How often a field is found when it exists in the document
F1 Score — Combined measure of precision and recall

Predictions & Review

After ReadAI extracts data from a document, the result is called a Prediction. The Predictions view lets you review, correct, and approve extracted data.

Viewing Predictions

Go to ReadAI and click the Predictions tab to see all processed documents and their extraction results.

Each row shows:

Document name and type
Processing date
Extraction model used
Overall confidence score
Status (Pending Review / Approved / Rejected)

Reviewing a Prediction

Click any prediction to open the side-by-side review view:

Left panel: The original document (rendered)
Right panel: All extracted fields with their values and confidence scores

Approving

If all extracted values are correct, click Approve. The data is marked as verified and can be sent downstream.

Correcting

If a field value is wrong:

Click the field value.
Type the correct value.
Click Save Correction.

Corrected predictions are flagged and can be used for model retraining.

Rejecting

If the document is unreadable or the extraction is severely incorrect, click Reject and provide a reason. The document is excluded from downstream processing.

Bulk Review

For high-volume workflows:

Filter by status = “Pending Review”.
Select multiple predictions.
Click Bulk Approve to approve all selected (use when confidence is uniformly high).