Explaining a Single Problem: Far More Expensive Than You'd Imagine

I. The Business Pain Point: Explaining Problems Is Unbelievably Expensive

We are a five-person team building a K12 math practice app. Three months after launch, the most requested feature from user feedback wasn't more practice problems—it was "wrong-answer explanations." When students get a question wrong, they want to know why they got it wrong.

Initially, we had two part-time teachers writing explanations manually, and we quickly ran the numbers:

With ~2,000 wrong answers per day, the manual model cost over 10,000 RMB daily—completely unviable. The conclusion was clear: wrong-answer explanations had to be AI-powered, but there's an engineering gap between "works in a demo" and "ready for production."

II. Architecture Design: A Three-Layer Structure Where Explanation Quality Is Guaranteed by Process

Many people think wrong-answer explanation is just "throwing the question at a large model." In practice, direct API calls had three problems: messy formula rendering, explanation depth mismatched with the student's grade level, and maintaining multiple model APIs separately was exhausting. We ultimately landed on a three-layer architecture:

┌─────────────────────────────────────┐
│  应用层：拍照/输入题目 → 用户学段/教材版本  │
├─────────────────────────────────────┤
│  业务编排层（核心）                       │
│  1. 预处理：OCR识别 + LaTeX公式还原校验    │
│  2. 知识点标注：轻量模型打标（小模型，便宜）  │
│  3. 讲解生成：强模型 + 分学段提示词模板      │
│  4. 质量校验：答案复核 + 步骤数检查         │
│  5. 失败降级：换模型重试 / 转人工队列       │
├─────────────────────────────────────┤
│  AI API网关层（统一ThisToken.AI接入）      │
│  多模型统一调用 · 统一计费 · 统一监控       │
└─────────────────────────────────────┘

Key decision: the code never depends directly on any vendor's official SDK—all model calls go through the unified gateway using OpenAI-compatible formats.

III. Key Implementation Steps

Step 1: Break down tasks and assign models by capability. Knowledge point tagging uses a cheap small model, explanation generation uses a strong reasoning model, and answer verification uses a mid-tier model. Overall calling costs dropped ~70% compared to "using the flagship model for everything."

Step 2: Grade-specific prompt templates. The tone and step density appropriate for a fifth grader versus a ninth grader are completely different. We maintain 6 template sets, injected based on user profiles.

Step 3: Quality gate. After an explanation is generated, an answer consistency check is mandatory—the final answer in the AI's explanation must match the question's standard answer. Any mismatch immediately triggers a model-switch retry; if it still fails, it goes to the human queue. This is the bottom line for user trust.

Core calling code example:

from openai import OpenAI

client = OpenAI(
    api_key="your-thistoken-key",
    base_url="https://api.thistoken.ai/v1"  # 统一网关
)

def explain_question(question, grade, kp_tags):
    prompt = PROMPT_TEMPLATES[grade].format(
        question=question, knowledge_points=kp_tags
    )
    resp = client.chat.completions.create(
        model="deepseek-chat",          # 讲解生成用性价比模型
        messages=[{"role": "user", "content": prompt}],
        temperature=0.3
    )
    return resp.choices[0].message.content

def verify_answer(question, explanation, answer):
    # 独立模型复核，防讲解"讲歪"
    check = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content":
            f"判断该讲解最终答案是否等于{answer}：{explanation}"}]
    )
    return "一致" in check.choices[0].message.content

Post-launch comparison:

IV. Why a Unified AI Gateway Reduces Maintenance Costs

We took some detours before launch: early on, we integrated a vendor's official SDK directly, and within three months encountered two model version changes and one pricing change. Each required code changes and regression testing, costing the team roughly 2-3 person-days per adaptation. After switching to a unified gateway:

1. One SDK handles all models. With OpenAI-compatible format access, switching the explanation or verification model means changing a single model string—no second dependency needed. We trialed 4 models before finding the best cost-performance combination, without touching the architecture once.
2. No registering and reconciling accounts with each vendor separately. Bills for five models become one; cost monitoring has a unified interface. We split call statistics by feature, and monthly reconciliation went from half a day down to ten minutes.
3. Vendor risk is isolated. If a model changes pricing or gets rate-limited, the fallback logic simply swaps models at the gateway level with zero changes to business code—this saved us twice when traffic doubled during peak education season.

For a five-person team, the saved adaptation work is equivalent to one extra person's development time per month. That's the real value of a unified gateway: it's not about saving money, it's about saving the most expensive resource—engineers' attention.

V. Final Thoughts

After launch, the wrong-answer explanation feature became the module with the longest in-app session time, while its marginal cost is nearly negligible. The feasibility formula for AI education apps is actually simple: thorough task decomposition + a quality gate as a safety net + a unified gateway to control costs.

If you want to quickly validate your own AI feature ideas, you can start by registering an account on ThisToken.AI to access multiple models through a single integration. From your first line of calling code to seeing results might take just ten minutes: https://api.thistoken.ai/register

---

Ready to try it yourself? Sign up at https://api.thistoken.ai/register to get your API key and start building.