Knowledge Graph#

The Knowledge Node List#

Each row in the list is a knowledge node — one piece of accounting guidance the AI uses. The table has the following columns:

• Code — a short identifier for the node (e.g. VENDOR-GRAB-DR).
• Title — a plain-language name for the rule.
• Summary — a brief description of what the rule covers.
• Type — the kind of node: Rule, Guideline, Vendor, Pattern, or Learned. "Learned" nodes are created automatically when the AI observes a coding pattern from your corrections.
• Category — an optional grouping label you assign.
• Enabled — whether the AI is currently consulting this node.
• Reliability — how confident and well-used the node is:
  • If the node has never been used by the AI, this column shows a dash.
  • If the node has been used, it shows a confidence percentage and a usage count (for example 82% · 14×). Green means high confidence (70% or above), amber means moderate (40–69%), red means low (below 40%).
  • If the node has been retired, it shows a Retired badge instead.

Learning Health Card#

When the AI has processed documents and received corrections, a Learning health summary card appears above the table. It shows:

• Corrections — the total number of times an operator has corrected an AI-generated journal entry, which the system used to learn new vendor coding patterns.
• Documents processed — how many source documents the AI has extracted in total.
• Correction rate — the proportion of processed documents that required a correction. A falling rate over time means the AI is improving.
• Most-corrected vendors — up to three vendors whose documents most often needed a correction, so you know where to focus review.

This card only appears once the system has enough data. If it is not visible, the AI has not yet accumulated corrections for your company.

Coding Conflicts Card#

If the AI has learned contradictory coding patterns for the same vendor — for example, seeing the same vendor debited to two different accounts in different documents — a Coding conflicts to resolve card appears above the table. Each row in the card names the vendor and shows the competing account codes and how often each appeared.

To resolve a conflict, find the relevant Learned node in the list below and either edit it to reflect the correct account, retire it (see below), or leave the conflicting entries for a finance reviewer to decide.

Adding a Node#

1. Click Add Knowledge Node (top right of the page).
2. Fill in the form:
   • Type (required) — choose the kind of node.
   • Title (required) — a short, clear name.
   • Summary (required) — one or two sentences explaining the rule.
   • Body — optional longer explanation or examples.
   • Category — optional grouping label.
   • Enabled — leave checked to make the node active immediately.
3. Click Save.

Editing a Node#

1. Click the row for the node you want to change.
2. Update the fields as needed.
3. When editing an existing node, a Learning status field also appears with three options:
   • Active — the node is live and the AI consults it.
   • Candidate — the node is under review; the AI still consults it but it is flagged for human verification.
   • Retired — the node is disabled. The AI will no longer use it, and its Reliability column shows a Retired badge. Use this when the AI has learned a wrong pattern (for example an incorrect account code) and you want to stop it from being applied to future documents.
4. Click Save.

To reactivate a retired node, edit it and set the Learning status back to Active.

Deleting a Node#

Click the delete icon on the node's row and confirm. Deleting is permanent. If you are unsure, use Retired status instead — you can always reactivate a retired node, but a deleted node is gone.

AI Confidence on the Documents List#

The Treasury → Accounting → Documents list has an AI Confidence column. This shows how confident the AI was when it extracted account codes from each source document:

• Green badge (N%) — confidence is 70% or above; the extraction is likely accurate.
• Amber badge (N%) — confidence is 50–69%; review the journal entry before posting.
• Red "Review · N%" badge — confidence is below 50%; the AI was uncertain. Open the document and check the suggested accounts against your chart of accounts before posting.

A document with no badge was not processed by the AI (it was entered manually or extracted before confidence scoring was added).

Common Tasks#