AI 扩展计划 / Playbooks

ABPA 模板 08：AI Architecture ADR Set

145 行abpa/templates/08-ai-architecture-adr-set.md

AI Architecture ADR Set

Use this template to document architecture choices for an AI product. Each ADR should be short, specific, and reversible when evidence changes.

ADR Index

ADR	Topic	Status	Decision date	Owner	Revisit trigger
ADR-001	AI pattern selection	proposed / accepted / superseded
ADR-002	Model and provider strategy	proposed / accepted / superseded
ADR-003	RAG and knowledge architecture	proposed / accepted / superseded
ADR-004	Human-in-the-loop design	proposed / accepted / superseded
ADR-005	Eval and observability architecture	proposed / accepted / superseded
ADR-006	Integration, security, and audit	proposed / accepted / superseded

ADR-001: AI Pattern Selection

Field	Answer
Context
Business capability
Workflow step
Decision needed	choose workflow automation / RAG assistant / copilot / agent / rules-only

Options

Option	Fit	Strength	Weakness	Evidence needed
Rules-only	low / medium / high
Search / RAG assistant	low / medium / high
Copilot with user approval	low / medium / high
Bounded agent	low / medium / high
Fully automated decision	low / medium / high

Decision

Field	Answer
Selected pattern
Why
Consequence
Reversal trigger

ADR-002: Model And Provider Strategy

Field	Answer
Primary model
Fallback model
Hosting pattern	API / managed cloud / self-hosted / hybrid
Data sensitivity constraint
Cost constraint

Criterion	Weight	Option A	Option B	Option C
Task quality
Latency
Cost per case
Data control
Operational maturity
Vendor risk

ADR-003: RAG And Knowledge Architecture

Field	Answer
Retrieval purpose	explain / summarize / classify / recommend / draft
Knowledge sources
Index owner
Refresh cadence
Citation requirement

Design choice	Selected option	Reason	Risk
Chunk strategy
Metadata model
Retrieval method	keyword / vector / hybrid / graph
Reranking	yes / no
Grounding requirement	citation / quote / source id / evidence span
Stale content handling

ADR-004: Human-In-The-Loop Design

Field	Answer
Human role	reviewer / approver / exception handler / trainer
Required approval points
Escalation owner
Override logging

AI action	Autonomy level	Human checkpoint	Stop condition
Retrieve evidence	suggest / execute / decide
Draft summary	suggest / execute / decide
Recommend action	suggest / execute / decide
Send external output	suggest / execute / decide

ADR-005: Eval And Observability Architecture

Field	Answer
Primary eval dataset
Golden labels owner
Release gate
Production monitoring owner

Signal	Purpose	Collection point	Threshold	Response
Task success	quality	offline / shadow / prod
Hallucination / unsupported claim	safety	offline / shadow / prod
Latency	UX	prod
Cost per case	ROI	prod
User override rate	trust	prod
Incident rate	risk	prod

ADR-006: Integration, Security, And Audit

Field	Answer
Systems integrated
Auth pattern
Secrets handling
Audit requirement
Data retention

Risk	Architecture control	Evidence
Unauthorized data access
Prompt injection
Sensitive output leakage
Tool misuse
Missing audit trail
Vendor outage

Architecture Decision Summary

Question	Answer
What are we building?
What are we deliberately not building?
What evidence supports this design?
What could make this decision wrong?
What should be tested next?