LLM Apps in Production (RAG + Vector DB + Caching) · 课时

为任务选择合适的模型

学习如何使用模型分层、级联和质量门控，将每个请求路由到能够高质量完成任务的最低成本模型，从而降低 RAG 的成本和延迟。

为任务选择合适的模型 是 CoddyKit 上的免费 LLM Apps in Production (RAG + Vector DB + Caching) 课时。这是第 4 节课，共 4 节。你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。这是 LLM Apps in Production (RAG + Vector DB + Caching) 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 LLM Apps in Production (RAG + Vector DB + Caching) 课程共包含 4 节课。

本课时的部分内容尚未翻译，以英文显示。

Why Model Choice Drives Cost

In a RAG pipeline the LLM call is usually the single biggest cost and latency driver. The same prompt sent to a flagship model can cost 20-50x more than a small model.

Optimizing model selection is often the highest-leverage change you can make.

Token price differs per model
Latency scales with model size
Not every query needs the biggest brain

Model Tiers

Group your available models into tiers by capability and price:

Small / cheap — classification, extraction, simple Q&A
Mid — most RAG answers grounded in retrieved context
Large / flagship — multi-step reasoning, ambiguous queries

Default to the smallest tier that meets your quality bar.

A Simple Router

A router inspects the request and picks a model. Start with rule-based routing before adding ML.

def pick_model(query, context_len):
    if len(query) < 80 and context_len < 2000:
        return 'small-model'
    if 'explain' in query or 'compare' in query:
        return 'large-model'
    return 'mid-model'

print(pick_model('What is the price?', 500))

Model Cascades

A cascade tries a cheap model first, then escalates only if the answer is low confidence. Most queries resolve cheaply; only the hard ones reach the expensive model.

Run small model
Score confidence / check guardrails
Escalate only on failure

Cascade in Code

A minimal cascade with a confidence check.

def answer(query):
    cheap = call('small-model', query)
    if cheap['confidence'] >= 0.8:
        return cheap['text']
    return call('large-model', query)['text']

def call(model, query):
    return {'text': 'stub', 'confidence': 0.9}

print(answer('hello'))

Confidence Signals

How do you know the cheap answer is good enough? Useful signals:

Self-reported confidence from the model
Whether the answer cites retrieved context
Output length / refusal patterns
A small judge model scoring the answer

Matching Context Size to Model

Large context windows are expensive. A model that accepts 200k tokens charges you for every token you send. Trim retrieved chunks aggressively and reserve big windows for queries that truly need them.

Right-sizing context is part of right-sizing the model.

Measuring Quality per Tier

Before downgrading a model, measure quality on a fixed eval set. Track accuracy per tier so you know the real trade-off.

scores = {'small': 0.81, 'mid': 0.90, 'large': 0.93}
bar = 0.88
cheapest_ok = next(m for m, s in scores.items() if s >= bar)
print('Use:', cheapest_ok)

Cost vs Quality Curve

Plotting cost against quality usually shows diminishing returns: jumping to the flagship model buys a few points of accuracy at multiples of the cost.

Pick the point where quality crosses your acceptance bar at the lowest cost.

Fallbacks for Reliability

Routing also helps reliability. If your primary model is rate-limited or down, route to an alternative provider of similar tier so users still get answers.

Primary -> secondary provider
Same tier, comparable quality
Log which path served the request

Putting It Together

A production router combines: tier rules, a cascade for hard queries, context trimming, and provider fallbacks. Continuously evaluate so routing stays calibrated as models change.

Quick Check

Test your understanding of model cascades.

Recap

You learned to cut RAG cost and latency by choosing the right model: tier your models, default to the smallest that meets your bar, use cascades to escalate only hard queries, right-size context, and keep provider fallbacks for reliability. Always validate routing against an eval set.

常见问题解答

「为任务选择合适的模型」课时是免费的吗？

是的 — 「为任务选择合适的模型」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 LLM Apps in Production (RAG + Vector DB + Caching) 课程的其余内容，请升级到 CoddyKit PRO。 LLM Apps in Production (RAG + Vector DB + Caching) 课程共包含 4 节课。

「为任务选择合适的模型」这节课中我会学到什么？

学习如何使用模型分层、级联和质量门控，将每个请求路由到能够高质量完成任务的最低成本模型，从而降低 RAG 的成本和延迟。你通过在浏览器中直接运行的动手代码来练习 LLM Apps in Production (RAG + Vector DB + Caching)，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 LLM Apps in Production (RAG + Vector DB + Caching) 需要有经验吗？

无需任何先前经验。CoddyKit 上的 LLM Apps in Production (RAG + Vector DB + Caching) 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。这是第 4 节课，共 4 节。

「为任务选择合适的模型」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 LLM Apps in Production (RAG + Vector DB + Caching) 课中编写并运行代码吗？

能。每节 LLM Apps in Production (RAG + Vector DB + Caching) 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

← 返回 LLM Apps in Production (RAG + Vector DB + Caching)