RAG in Wealth Management — Fiduciary Constraints and Retrieval Design

4.1 Domain Characteristics and Challenges

Wealth management AI operates under the most severe fiduciary and regulatory constraints of any financial domain. The system is simultaneously an analyst, a compliance officer, a tax advisor, and a relationship manager — but is legally constrained from acting as any of them without human oversight.

Regulatory landscape:

• Canada: IIROC (now CIRO), OSC, OSFI, FATF (AML)
• US: FINRA, SEC Rule 15l-1 (Reg BI — Best Interest)
• EU: MiFID II (suitability, appropriateness)
• FATF: AML/KYC — beneficial ownership, PEP screening

Domain challenges:

4.2 Level 1 — Vanilla RAG

Use case: Product knowledge base — "What is a GIC?", "How does a TFSA contribution room work?", "What is a dividend reinvestment plan?"

Implementation: Embed product brochures, regulatory fact sheets, knowledge articles → semantic search → FAQ response generation.

Real-world example: RBC's MyAdvisor chatbot handles basic product FAQs. TD's EasyWeb virtual assistant deflects tier-1 queries. These are all effectively L1.

What L1 handles well:

• Product education content (low-risk, no personalization needed)
• General tax-advantaged account explanations (RRSP, TFSA, FHSA)
• Process FAQs (how to transfer a RRSP, how to change beneficiaries)

Where L1 dangerously fails:

• "Should I put my RRSP contribution into Canadian equities or bonds given current market conditions?" — this is a suitability question. An L1 system that answers this without client profile context is making an unsuitable recommendation and is in regulatory violation.
• L1 has no mechanism to know it doesn't have enough context to answer.

4.3 Level 2 — Hybrid RAG

Use case: Investment research retrieval, earnings report analysis, portfolio news scanning.

Why hybrid matters for financial text:

• Analyst reports are dense with exact ticker symbols, CUSIP identifiers, financial ratios — BM25 is essential for exact-match on these
• But the semantic intent behind "show me bearish analyst coverage on Canadian bank stocks" requires dense retrieval
• Financial jargon evolves: "stealth tightening", "quantitative tightening", "soft landing" — dense vectors capture semantic drift better

Concrete example — Portfolio research assistant:

Query: "What are analysts saying about the rate sensitivity of Canadian bank NIM in 2025?"

Dense path: semantic similarity to "net interest margin rate sensitivity Canadian banks 2025"
            → retrieves analyst commentary, earnings call transcripts, sector reports

Sparse path: BM25 exact match on ["NIM", "net interest margin", "TD", "RY", "BNS", "BMO", "CM"]
            → retrieves reports with specific bank identifiers and exact financial metric names

Reranker:  Cross-encoder scores and prioritizes docs that contain BOTH semantic context
           AND exact bank ticker identifiers

Result: 5 analyst reports specifically discussing Canadian bank NIM in rate-sensitive scenarios

Real-world systems:

• Bloomberg Terminal BQNT — BQuant combines NLP semantic search with structured data exact queries
• FactSet Research Systems — hybrid retrieval over fundamental data + document corpus
• Refinitiv Eikon — semantic document search + exact financial identifier matching
• Morgan Stanley's AI @ Morgan Stanley (2023, OpenAI-powered) — assistant retrieves from 100k+ research reports using hybrid search

4.4 Level 3 — GraphRAG

Use case: Regulatory suitability assessment, conflict of interest screening, portfolio constraint compliance.

Wealth management ontology:

Query example — Suitability check:

"Is this 7-year autocallable note on an equity basket suitable for this client?"

Graph traversal:

1. Client → InvestorType: Retail (not accredited) → limits product universe
2. Product → [REQUIRES_CATEGORY: Accredited] → IMMEDIATELY PROHIBITED for retail
   [If accredited: continue]
3. Product → RiskRating: High → Client RiskProfile: Conservative → MISMATCH
4. Client → IPS → [CONSTRAINT: max_alternative: 10%] → current alternatives allocation = 8%
5. Product → LiquidityRating: Illiquid (7-year lockup) → Client [LIQUIDITY_NEED: High] → MISMATCH
6. Product → [SUBJECT_TO: OSC Rule 45-106] → offering memorandum required
7. Synthesis: 2 hard stops (investor type + risk mismatch), 1 soft concern (liquidity)
8. Generate: Unsuitable — cannot recommend; document rationale for compliance record

Real-world examples:

• SuitabilityPro (Broadridge) — uses rule-based graph engines for suitability; moving toward graph + LLM hybrid
• Compliance.ai — regulatory change graph that maps regulatory updates to affected products and client segments
• Backstop Solutions (BlackRock) — portfolio constraint compliance using graph-based rule engines

4.5 Level 4 — Agentic RAG

Use case: Proactive portfolio review triggered by market events, tax-loss harvesting opportunity identification, next-best-action for advisors.

Example — IPS drift detection:

"Given Q3 market events, identify which of our high-net-worth clients have portfolios that have drifted beyond their IPS constraints and require rebalancing."

Agent loop:

Turn 1: Query market events Q3 2024 → identify major themes: rate cuts, CAD weakness, energy sector decline
Turn 2: Identify affected asset classes and sectors → Canadian equities (-4%), energy (-12%), REITs (+3%)
Turn 3: For each HNW client: retrieve portfolio holdings
Turn 4: For each client: retrieve IPS constraints (max equity %, sector exclusions, benchmark)
Turn 5: Calculate current allocation vs. IPS target for each client
Turn 6: Reflect: Which clients exceed tolerance bands (typically ±5% from target)?
Turn 7: For flagged clients: retrieve rebalancing options within IPS constraints
Turn 8: Rank by urgency and tax impact (RRSP vs. non-registered rebalancing has different tax treatment)
Turn 9: Generate advisor briefing: client list + drift magnitude + recommended trades + tax impact estimate

Real-world systems:

• Morgan Stanley Next Best Action (NBA) — advisor-facing LLM that surfaces proactive client opportunities
• Salesforce Einstein for Wealth Management — CRM + portfolio data agentic retrieval
• RBC's Aiden (internal) — AI model used for client segmentation and advisor recommendations
• Vanguard's Personal Advisor Services — semi-agentic portfolio review and rebalancing triggers

4.6 Supporting Elements — Wealth Management Domain

Memory:

Client relationship memory:
  - Investment policy statement (IPS) — evolves over time; version history matters
  - Life event flags (retirement approaching, inheritance received, divorce)
  - Communication preferences and meeting history
  - Prior recommendations made and client responses

Market context memory:
  - Rolling 90-day market event log
  - Active regulatory changes affecting product universe
  - Interest rate and currency regime context

Compliance memory:
  - Rationale log for every recommendation (audit trail — regulatory requirement)
  - Suitability assessment history per product per client
  - Complaints and dispute records

Prompt Engineering:

Fiduciary system prompt:
  "You are a portfolio analytics assistant supporting a registered investment advisor.
   You do not give investment advice directly to clients. You support advisor decision-making.
   Client profile: [INJECT_CLIENT_SUMMARY]
   IPS constraints: [INJECT_IPS]
   Current portfolio: [INJECT_HOLDINGS_SNAPSHOT with TIMESTAMP]
   Regulatory context: [INJECT_ACTIVE_REGULATIONS for client.jurisdiction]
   Required: Always distinguish between (a) factual portfolio analysis and (b) forward-looking recommendations.
   Required: Every recommendation must cite IPS section and regulatory basis.
   Required: Flag any recommendation requiring compliance pre-approval."

Constraint-awareness prompting:
  "Before answering, identify: What does this client's IPS prohibit? What investor category are they?
   What is their stated risk tolerance? Answer only within those guardrails."

Fine-Tuning:

• FinBERT: Pre-trained on financial news, SEC filings, earnings transcripts — better semantic alignment for financial text
• Instruction fine-tuning on earnings Q&A: Teach the model to extract net interest margin, CET1 ratio, efficiency ratio from dense earnings transcripts without confusing similar metrics
• Compliance fine-tuning: Train the model on thousands of suitability determination examples so it learns the decision boundary (suitable vs. unsuitable) with high precision
• Tool: Unsloth + LoRA on LLaMA-3 base; domain-adapted FinBERT as the embedding backbone

Part of the RAG Enterprise Series. Next: RAG in Personal Banking — Scale, AML, and Transaction Intelligence.