규제 도메인을 위한 NIST AI RMF 준수 RAG 시스템을 구축한 방법

DEV Community · Lakshman Pandey · 2026-08-18 개발(SW)

How I Built a NIST AI RMF-Compliant Production RAG System

By Lakshman Pandey | August 2026

Introduction

I shipped a production RAG (retrieval-augmented generation) system serving UK arts and culture clients. This article documents how the system implements NIST AI Risk Management Framework controls, with real decisions, trade-offs, and measurable outcomes.

TL;DR:

Designed for low-risk retrieval-grounded use cases
All decisions documented in ADRs (Architecture Decision Records)
Costs $0.003-0.005 per query, $25/month infrastructure
EU data residency (GDPR-ready)
Eval framework prevents quality degradation

The System

Stack:

Frontend: Streamlit Cloud
Vector DB: Supabase pgvector (EU-West-2)
Embeddings: Voyage AI (1024 dimensions)
LLM: Claude Haiku 4.5 (direct REST API)
Observability: Langfuse
Integration: MCP server for Claude Desktop

Risk Profile: LOW

Retrieval-grounded (not generative by itself)
No model training or fine-tuning
Human review possible before deployment
No safety-critical decisions

NIST AI RMF Implementation

The NIST framework has 4 functions: GOVERN, MAP, MEASURE, MANAGE. Here’s how the production system implements each.

1. GOVERN: Establishing Governance Structure

Requirement: Define roles, responsibilities, and decision-making authority for AI risk management.

Implementation:

Decision Authority: Solo architect with client stakeholder approval loops.

Data Governance (ADR-001):

Decided: Use Supabase pgvector in EU-West-2 (London)
Why: UK public-sector cultural clients require UK/EU data residency for GDPR compliance
Risk: Vendor dependency on Supabase
Mitigation: Eval framework + ADR ensures reversibility

Stakeholder Roles:

Developer: Lakshman (me) — system architecture, data pipeline, deployment
Client: UK public-sector stakeholder — approve governance policies, validate output quality
Operations: Future (TBD) — monitoring, alert response

Policy: All user data stays in EU. API calls to Claude/Voyage are transient (no data stored in US).

Measurement: Langfuse audit trail logs every query’s origin and destination.

2. MAP: Identify AI Risks

Requirement: Identify risks specific to the AI system’s context, design, and use case.

Implementation:

Risk Inventory:

Risk Severity Source Mitigation Hallucination Medium LLM generating answers beyond retrieved context Prompt constraints (answer only from sources) + eval suite thresholds Embedding Quality Drift Medium Voyage model updates degrade retrieval Phase 2 eval baseline (context recall 0.98) prevents regression Data Drift Low Corpus content changes over time Scheduled re-eval (monthly) against golden questions Vendor Outage Low Supabase/Voyage API downtime Documented fallback to Ollama (local, offline) PII Leakage Low User data in prompts (Future) Microsoft Presidio redaction at ingestion Prompt Injection Low User query attempts to jailbreak system Input validation + output validation (present but basic)

Risk Rating: OVERALL = LOW-RISK

Retrieval-grounded (not generative-primary)
Small, controlled corpus (108 documents)
Limited stakeholders (internal + client team)
No real-time safety-critical decisions

3. MEASURE: Establish Metrics & Monitoring

Requirement: Define metrics to assess AI system performance and risk throughout the lifecycle.

Implementation:

Eval Framework (Phase 2):

Built Ragas-based evaluation suite with 18 golden questions:

Metric Baseline Threshold Current Status Faithfulness 0.42 > 0.50 Pending re-run with Voyage Context Precision 0.69 > 0.65 ✅ Passing Context Recall 0.98 > 0.95 ✅ Passing Answer Relevancy 0.64 > 0.60 ✅ Passing

Why these metrics?

Faithfulness: Detects hallucination (answers consistent with sources)
Context Precision: Ensures retrieved chunks are actually relevant
Context Recall: Ensures good chunks aren’t missed
Answer Relevancy: Ensures answer matches the question

Observability (Phase 3):

Langfuse integration traces every production query:

{
  "trace_name": "rag_query",
  "input": "What are the current content guidelines?",
  "output": "Based on sources [1][2]...",
  "input_tokens": 450,
  "output_tokens": 85,
  "cost_usd": 0.0031,
  "latency_ms": 1250
}

Enter fullscreen mode Exit fullscreen mode

Cost Per Query: $0.0005 (embedding) + $0.003 (generation) = $0.0031

Latency Target: < 2 seconds (currently ~1.2s)

4. MANAGE: Implement Risk Mitigation

Requirement: Manage identified risks through safeguards, monitoring, and response.

Implementation:

Current Safeguards:

Prompt Engineering: System prompt enforces “answer only from sources” constraint
Error Handling: Try-catch blocks prevent crashes; errors logged to Langfuse
Rate Limiting: (Future) Add max queries/hour per session
Cost Ceiling: (Future) Hard cap on monthly spend per client

Code Evidence:

# phase3-deployment/app.py, lines 52-58
response = requests.post(
    "https://api.anthropic.com/v1/messages",
    json={
        "model": "claude-haiku-4-5-20251001",
        "messages": [{
            "role": "user",
            "content": (
                'Answer using ONLY the sources below. '
                f'If answer not in sources, say so.\n\n{context}\n\nQ: {question}'
            )
        }]
    }
)

Enter fullscreen mode Exit fullscreen mode

Future Safeguards (Phase 4):

Human-in-the-loop approval gate for sensitive queries
Prompt caching to reduce costs by 25-50%
Microsoft Presidio for PII redaction
CI/CD regression gate (eval suite must pass before deploy)

Real-World Trade-offs

ADR-001: Supabase vs Self-Hosted PostgreSQL

Decision: Cloud-managed Supabase pgvector (EU)

Trade-off:

Gain: Managed backups, EU residency, zero DevOps
Cost: Vendor lock-in, moderate migration cost if Supabase changes

Why this trade-off wins:

Team of one (no DevOps capacity)
Clients demand EU data residency
Long-term value of EU compliance > switching cost

ADR-002: Voyage AI vs Ollama

Decision: Cloud API (Voyage) vs local (Ollama)

Trade-off:

Gain: Cloud-native, 1024 dims (better quality), managed updates
Cost: $0.0001 per embedding, vendor dependency

Why this trade-off wins:

Scales to 10 clients without infrastructure changes
Quality improvement (1024 vs 768 dims) is measurable
Cost per query is sub-penny

ADR-003: Direct REST Calls vs Anthropic SDK

Decision: Manual HTTP calls (requests lib) vs SDK

Trade-off:

Gain: Works on Python 3.14, fewer dependencies, explicit control
Cost: No type hints, manual error handling

Why this trade-off wins:

Python 3.14 breaks SDKs (httpx/httpcore incompatibility)
Direct API calls = future-proof
Explicit contract = easier to debug

Measuring Against NIST

GOVERN: ✅ Documented roles, EU data residency, stakeholder approval

MAP: ✅ Risk inventory, low-risk classification, identified mitigations

MEASURE: ✅ Eval framework (Phase 2), Langfuse tracing (Phase 3), cost monitoring

MANAGE: ⚠️ Basic error handling, prompt constraints; future human-in-the-loop + spend ceiling

Compliance Status: COMPLIANT with NIST for low-risk use case. Future enhancements (Phase 4) will strengthen MANAGE function.

Lessons Learned

Data Residency First: For UK public-sector clients, EU hosting is table-stakes. Chose Supabase before other factors.
Evaluate Everything: Phase 2 eval framework caught that naive keyword-matching underperforms vector search. Measuring > assuming.
Direct API > SDKs for Stability: Python 3.14 broke 4 versions of the Anthropic SDK. Direct HTTP calls worked immediately.
Cost Transparency Builds Trust: Langfuse tracing makes per-query costs visible. Clients appreciate this.
Document Decisions, Not Just Code: ADRs explain WHY, not just HOW. Critical for onboarding + architectural clarity.

What’s Next (Phase 4)

Human-in-the-loop approval for high-risk queries
Prompt caching (25-50% cost savings)
PII redaction at ingestion (Microsoft Presidio)
CI/CD regression gates (eval must pass)
Re-run Phase 2 evals with Voyage embeddings (prove quality parity)

References

ADR-001: Supabase pgvector (EU data residency)
ADR-002: Voyage AI embeddings (production-grade, cloud-native)
ADR-003: Direct REST API (Python 3.14 stability)
Phase 2 Eval: Ragas framework + 18 golden questions
Phase 3 System: Live production deployment
NIST AI RMF: https://www.nist.gov/itl/ai-risk-management-framework
Supabase pgvector: https://supabase.com/docs/guides/database/extensions/pgvector
Voyage AI: https://docs.voyageai.com/docs/embeddings
Claude API: https://docs.anthropic.com/en/api/getting-started

About

Lakshman Pandey is a Senior Technical Lead specializing in AI Solutions Architecture for content-rich, regulated domains (UK public sector, cultural institutions, education). 13+ years full-stack development (Drupal, Python, Node.js). Currently building RAG systems that balance innovation with governance requirements.

GitHub: code-lakshman/ai

원문에서 계속 ↗

추출 본문 · 출처: dev.to · https://dev.to/lakshman-ai/how-i-built-a-nist-ai-rmf-compliant-rag-system-for-regulated-domains-jch

규제 도메인을 위한 NIST AI RMF 준수 RAG 시스템을 구축한 방법

How I Built a NIST AI RMF-Compliant Production RAG System

Introduction

The System

NIST AI RMF Implementation

1. GOVERN: Establishing Governance Structure

2. MAP: Identify AI Risks

3. MEASURE: Establish Metrics & Monitoring

4. MANAGE: Implement Risk Mitigation

Real-World Trade-offs

ADR-001: Supabase vs Self-Hosted PostgreSQL

ADR-002: Voyage AI vs Ollama

ADR-003: Direct REST Calls vs Anthropic SDK

Measuring Against NIST

Lessons Learned

What’s Next (Phase 4)

References

About

더 많은 게시물

에피소드 1: “세상에 나타난 작은 점”