RAG Ingestion Guide

This guide explains how to load the public AML knowledge base into a RAG system or local AML agent.

Purpose#

The wiki should be ingested as AML process knowledge, not as a source of final decisions. It can help a system explain AML concepts, prepare workflows, draft source plans and structure evidence. It should not be used to approve, reject or clear a subject without analyst review.

Recommended chunks#

Chunk pages by headings. Preserve:

• page title;
• section heading;
• source page URL;
• topic tags;
• boundary notes;
• tables and checklists;
• code-like process flows.

Loading order#

Source Register
→ Foundational Learning Path
→ AML Fundamentals
→ Operation & Source Workflow Logic
→ Open AML Agent Framework
→ Verification & Evidence Control Layer
→ S010 Stagegate Runner
→ LLM-to-S010 Bridge
→ LLM Orchestration Interface
→ Case File & Evidence Lifecycle
→ Discrepancy, Risk & Escalation Controls
→ MLRO Review & Escalation Overview

Retrieval patterns#

For a concept answer, retrieve the relevant AML fundamental page and the source register.

For a workflow answer, retrieve the operation workflow page, S010 control layer, case-file lifecycle and discrepancy/escalation controls.

For an LLM-agent implementation answer, retrieve Open AML Agent Framework, LLM-to-S010 Bridge, LLM Orchestration Interface, S010 Stagegate Runner and safety boundary sections.

Safety metadata#

Mark the following as persistent boundaries:

final_decision_allowed: false
analyst_review_required: true
llm_may_confirm_gates: false
llm_may_generate_gate_logs: false
llm_may_generate_auto_confirm_plan: false
llm_may_claim_live_screening_without_source_access: false

Local implementation data#

Do not embed private policies, credentials, client case files, domestic reporting forms, regulator portals or commercial database session data into the public wiki index. Those belong in a separate private implementation profile or secure local knowledge base.

Real-model validation#

If a local model is connected through an orchestration interface, test it repeatedly with the same prompt and inspect failed attempts, correction prompts and validated outputs. Real model output may include extra text, wrong enums, truncated JSON or invented fields. The system should treat those as validation events, not as trusted case output.