6 min read

Learning Engine

How Flux Capture learns from your corrections to improve accuracy

The Learning Engine is Flux Capture's adaptive intelligence system. Every correction you make helps the system get smarter, leading to increasingly accurate extractions over time.

How Learning Works

When you correct extracted data during review, Flux Capture remembers those corrections and applies them to future documents.

The Learning Loop

  1. Extract - AI extracts data from a document
  2. Review - You verify and correct any errors
  3. Learn - System stores your corrections
  4. Improve - Future extractions use learned patterns
  5. Repeat - Each correction improves accuracy

What the System Learns

Vendor Name Aliases

When the same vendor appears with different names on invoices:

Example:

  • Invoice says: "ACME Corporation Inc."
  • You select vendor: "ACME Corp"
  • System learns: "ACME Corporation Inc." → "ACME Corp"

Next time an invoice from "ACME Corporation Inc." arrives, the system automatically matches to "ACME Corp".

Date Format Patterns

Different vendors use different date formats:

Example:

  • Vendor A uses: MM/DD/YYYY (03/15/2024)
  • Vendor B uses: DD/MM/YYYY (15/03/2024)

When you correct a date interpretation, the system remembers that vendor's preferred format.

Account Coding Patterns

The system learns your GL account preferences:

Example:

  • Office supplies from Staples → Account 6200 (Office Supplies)
  • Shipping from UPS → Account 6100 (Shipping)

After a few corrections, the system suggests the right account automatically.

Amount Format Preferences

For vendors using different number formats:

Example:

  • European vendor uses: 1.234,56 (comma decimal)
  • US vendor uses: 1,234.56 (period decimal)

Building Vendor Intelligence

Over time, Flux Capture builds a profile for each vendor:

Vendor Defaults

Learned defaults include:

  • Payment terms
  • Default expense account
  • Typical invoice amounts
  • Expected date ranges
  • Currency preference

Invoice Patterns

The system tracks:

  • Invoice number formats
  • Typical line item descriptions
  • Average invoice amounts
  • Normal payment terms

✅ Tip: Process invoices from the same vendor consistently to build stronger vendor profiles.

Training the System

Making Effective Corrections

For best learning results:

  1. Always correct errors - Don't skip mistakes
  2. Be consistent - Use the same vendor names
  3. Complete the review - Don't abandon partial corrections
  4. Approve when ready - Approvals confirm learning

What Corrections Teach

Correction Type What System Learns
Vendor selection Name aliases, matching patterns
Date correction Date format preferences
Amount correction Number format patterns
Account selection Coding patterns by description
Currency change Currency preferences

Learning from Approvals

When you approve a document without changes, the system learns:

  • Extracted values were correct
  • Current patterns are working
  • Vendor matching was accurate

Learning Timeline

How quickly does accuracy improve?

Immediate Learning

Some patterns apply immediately:

  • Vendor name aliases
  • Explicit corrections

Pattern Recognition

Other patterns need multiple examples:

  • 3-5 corrections for date format recognition
  • 5-10 corrections for account coding patterns
  • 10+ documents for statistical patterns

Continuous Improvement

Over months of use:

  • Accuracy typically improves 10-20%
  • High-volume vendors improve fastest
  • New vendors start from baseline

Viewing Learned Data

Vendor Alias List

To see what the system has learned about vendor names:

  1. Go to Settings
  2. View the Learning section
  3. Browse vendor aliases

Account Suggestions

When reviewing a document:

  • Look for suggested accounts on line items
  • Suggestions appear based on learned patterns

Managing Learned Data

Reviewing Aliases

Periodically review learned aliases to ensure accuracy:

  1. Check that mappings are correct
  2. Remove any incorrect aliases
  3. Merge duplicate vendors

Resetting Learning

If learning has gone wrong for a vendor:

  1. Contact support to reset specific vendor learning
  2. Or clear all learning for a fresh start

⚠️ Warning: Resetting learning removes valuable accumulated intelligence. Only reset when necessary.

Best Practices

Consistent Vendor Names

Use a single canonical name for each vendor in NetSuite:

  • Good: "Office Depot" (always)
  • Bad: "Office Depot", "OfficeDepot", "Office Depot Inc."

Process Similar Invoices

Group invoices from the same vendor:

  • Helps build patterns faster
  • Reduces context switching

Correct Rather Than Skip

When extraction is wrong:

  • Correct the value (teaches the system)
  • Don't delete and re-upload (loses learning opportunity)

Complete Reviews

Finish what you start:

  • Partial reviews may save incomplete patterns
  • Complete reviews ensure accurate learning

How Learning Improves Accuracy

Before Learning

New installation, no learned data:

  • Vendor matching: ~70% accuracy
  • Account coding: Manual selection
  • Date parsing: Based on format detection

After 1 Month

With regular use:

  • Vendor matching: ~85% accuracy
  • Account coding: Suggestions for common vendors
  • Date parsing: Vendor-specific format awareness

After 6 Months

Mature installation:

  • Vendor matching: ~95% accuracy
  • Account coding: Auto-populated for most vendors
  • Date parsing: Rarely needs correction

Privacy and Data

What's Stored

The Learning Engine stores:

  • Vendor name mappings
  • Format preferences per vendor
  • Account coding patterns
  • Statistical patterns (no raw content)

What's Not Stored

  • Document content is not retained
  • Line item text is not stored long-term
  • Personal information is not tracked

Data Security

All learned data:

  • Stored in your NetSuite account
  • Not shared with other customers
  • Protected by NetSuite security

Troubleshooting Learning Issues

Not Learning Correctly

If the system isn't learning as expected:

  1. Verify you're completing reviews (not abandoning)
  2. Check that corrections are being saved
  3. Ensure vendors are being properly selected

Wrong Suggestions

If suggestions are consistently wrong:

  1. Make explicit corrections
  2. The new pattern will override old learning
  3. Contact support if issues persist

Slow Improvement

If accuracy isn't improving:

  1. Ensure consistent vendor usage
  2. Process more documents from problem vendors
  3. Review and correct all errors

Next Steps