行业知识 · v2.3.0 · 资料核对 2026-10-03

AI Solution Design

1. Decision and constraints

Product task and baseline:
Required quality / unacceptable failure:
Volume, latency, availability, and cost:
Data, privacy, region, and deployment:
Maximum autonomy and human oversight:

2. Option comparison

Option	Quality	Latency	Cost	Risk/control	Operations	Decision
Rule/search/traditional ML
Prompt-only
RAG
Workflow/tools
Bounded Agent
Fine-tuning

3. Proposed architecture

User / event
→ identity, permission, input validation
→ task routing and context construction
→ rules / retrieval / model / workflow / tools
→ output and domain validation
→ confirmation / action / response / handoff
→ trace, metrics, feedback, audit

Describe each component, owner, interface, data, timeout, retry, and failure behavior.

4. Data and context

Sources, owner, authority, freshness, ACL:
Ingestion, chunking/index, sync, deletion:
Context budget and ordering:
Memory scope and retention:
Untrusted-content isolation:

5. Models and routing

Task	Candidate	Reason	Fallback	Version / change policy

6. Tools and actions

Tool	Read/write	Authorization	Input validation	Idempotency	Confirmation	Recovery

7. Reliability and safety

Rate, time, step, token, and cost budgets:
Refusal / fallback / human handoff:
Tenant isolation and secret handling:
Audit and evidence:
Failure modes and controls:

8. Observability and versions

Trace fields and privacy treatment:
Model/Prompt/index/tool/code/eval versions:
Dashboards, alerts, sample review:

9. Evaluation and experiments

Hypothesis / decision	Dataset	Metric	Threshold	Owner	Result / next decision

10. Capacity and unit cost

Traffic assumptions:
P50/P95 tokens, steps, tool calls, latency:
Total cost per successful task:
Conservative/base/upside capacity:

11. Delivery and rollback

Prototype and phase gates:
Dependencies:
Rollout stages:
Rollback trigger/mechanism:
Open decisions and due dates: