Module 8: Your First AI System Design

4. Design Trade-Offs: Cost vs Latency vs Quality

Description

Reality: There's no "perfect" design. There are always trade-offs.

3 main dimensions:

  1. Cost: How much $ per request?
  2. Latency: How long does it take to respond?
  3. Quality: How good is the response?

The "trilemma": You can't maximize all 3 simultaneously.

  • Maximum quality (GPT-4) → high cost, high latency.
  • Minimum cost (GPT-3.5) → lower quality.
  • Minimum latency (cache) → static responses (lower personalized quality).

In this lesson you'll understand common trade-offs and how to make decisions.


Dimension 1: Cost

What Impacts Cost

Factors:

  1. The model: GPT-4 ($0.03/1K input) vs GPT-3.5 ($0.0005/1K input) → a 60× difference.
  2. Volume: 1K requests/month vs 1M requests/month.
  3. Prompt length: Long prompts (1K tokens) vs short ones (100 tokens).
  4. Response length: max_tokens=100 vs max_tokens=1000.
  5. The pattern: A simple chatbot (1 call) vs RAG (embeddings + an LLM call) vs Agents (multiple calls).

Trade-Off: GPT-4 vs GPT-3.5

Example: An FAQ chatbot

AspectGPT-4GPT-3.5
QualityVery high (superior reasoning)High (enough for an FAQ)
Cost/request$0.006 (200 input, 200 output)$0.0002 (30× cheaper)
Cost/month (10K requests)$60$2

Decision:

  • If it's a simple FAQ (direct answers) → GPT-3.5 (that's enough).
  • If it's a complex FAQ (reasoning) → GPT-4.

Trade-Off: Cloud API vs Self-Hosted

Example: 1M requests/month

AspectCloud (the GPT-4 API)Self-Hosted (Llama 3 70B)
Cost$6,000/month (variable)$1,500/month for the GPU (fixed)
QualityVery high (GPT-4)High (Llama 3 70B ~80% of GPT-4)
Setup5 minutes1-3 days

Decision:

  • If volume is low (<100K requests/month) → a cloud API (cheaper).
  • If volume is high (>1M requests/month) → self-hosted (4× cheaper).

Dimension 2: Latency

What Impacts Latency

Factors:

  1. The LLM: An API call (200-500ms) vs self-hosted locally (50-200ms).
  2. The pattern: A simple chatbot (1 call, <1s) vs RAG (embed + search + the LLM, 1-3s) vs Agents (multiple calls, 3-10s).
  3. max_tokens: 100 tokens (fast) vs 1000 tokens (slow).
  4. Cache: A hit (instant) vs a miss (an API call).
  5. The network: A US-based API call from LATAM (300ms) vs from the US (100ms).

Trade-Off: Quality vs Latency

Example: A RAG System

AspectRAG with GPT-4A simple chatbot with GPT-3.5
Latency2-3s (embed + search + the LLM)<1s (only the LLM)
QualityVery high (information specific to the docs)Medium (general information)

Decision:

  • If the user tolerates 2-3s and needs specific information → RAG.
  • If the user needs an instant response (<1s) and it's a general FAQ → a simple chatbot.

Trade-Off: GPT-4 vs Cache

Example: A repetitive FAQ

AspectGPT-4 (every request)Cache (a pre-generated response)
Latency500ms (an API call)<50ms (a Redis lookup)
QualityPersonalized (the user's context)Static (the same response every time)
Cost$0.006/request$0 (free after the first generation)

Decision:

  • If it's a frequent question ("What is AI?") and the answer doesn't change → cache.
  • If it's a personalized question ("How can I improve my code?") → GPT-4.

Dimension 3: Quality

What Impacts Quality

Factors:

  1. The model: GPT-4 (better reasoning) vs GPT-3.5 (enough for simple tasks).
  2. Prompt engineering: A clear system prompt vs a vague one.
  3. RAG: With docs (specific information) vs without docs (general information).
  4. Fine-tuning: A model adapted to a specific case vs a base model.
  5. Temperature: Low (deterministic) vs high (creative but inconsistent).

Trade-Off: Quality vs Cost

Example: A Ticket Classifier

AspectGPT-4GPT-3.5
Accuracy95%90%
Cost/request$0.002$0.00005 (40× cheaper)
Cost/month (100K tickets)$200$5

Decision:

  • If accuracy is critical (e.g. medical classification) → GPT-4.
  • If 90% accuracy is acceptable (e.g. support tickets) → GPT-3.5 (40× cheaper).

Trade-Off: Quality vs Latency (Agents)

Example: An Agent vs RAG

AspectAn Agent (multi-tool)RAG (a single retrieval)
QualityVery high (multi-step reasoning)High (single-step retrieval)
Latency5-10s (multiple calls)2-3s (1 call)
CostHigh (multiple LLM calls)Medium (1 LLM call)

Decision:

  • If the task is complex (multi-step reasoning) → an agent (the user tolerates 10s).
  • If the task is simple (Q&A) → RAG (2-3s is enough).

The Trilemma: You Can't Have Everything

Visualization:

        High Quality
              /\
             /  \
            /    \
           /      \
          /        \
   Low Cost ---- Low Latency

Reality:

  • High quality + low latency → high cost (GPT-4, no cache).
  • High quality + low cost → high latency (self-hosted Llama 3, not optimized).
  • Low latency + low cost → lower quality (a cache of static responses).

Optimization Strategies

1. A Multi-Model Strategy

Problem: Using GPT-4 for everything (expensive).

Solution:

  • GPT-3.5 for simple tasks (80% of requests) → $2/10K.
  • GPT-4 for complex tasks (20% of requests) → $12/10K.
  • Total: $14/10K (vs $60/10K with pure GPT-4 → a 75% saving).

2. Intelligent Caching

Problem: Repetitive FAQs burn tokens.

Solution:

  • The first time: GPT-4 generates the response → it's saved in the cache (Redis).
  • The second time (the same question): It returns from the cache (instant, free).
  • Saving: 90% of the cost on frequent questions.

3. Prompt Optimization

Problem: Long prompts (1K tokens) cost $.

Solution:

  • A concise prompt (200 tokens) → an 80% saving.
  • Few-shot examples (only 2-3, not 10).

4. Batch Processing

Problem: 1M individual requests (high overhead).

Solution:

  • Batch 100 requests into 1 call (e.g. for classification).
  • Saving: 50% of the cost (less overhead).

Real Use Cases

Case 1: A Startup MVP (a Limited Budget)

Requirements:

  • An FAQ chatbot.
  • Budget: $50/month.
  • Volume: 5K requests/month.

Decision:

  • Model: GPT-3.5-turbo (enough for an FAQ).
  • Cache: Redis for frequent questions (70% of requests).
  • Cost: $10/month (OpenAI) + $10/month (hosting) = $20/month.

Accepted trade-off: Quality slightly below GPT-4 (but 20× cheaper).


Case 2: A Company (High Volume, Critical Quality)

Requirements:

  • Q&A over technical documentation.
  • Volume: 500K requests/month.
  • Critical accuracy (>95%).

Decision:

  • Pattern: RAG (external docs).
  • Model: GPT-4 (accuracy >95%).
  • Vector DB: Pinecone (cloud, scalable).
  • Cache: Redis for frequent queries (30% of requests).
  • Cost: $3K/month (OpenAI) + $200/month (Pinecone) + $100/month (hosting) = $3.3K/month.

Accepted trade-off: A high cost (but accuracy is critical for the business).


Case 3: A Startup with Sensitive Data

Requirements:

  • An internal chatbot (confidential data).
  • Volume: 50K requests/month.
  • HIPAA compliance (data can't leave the infrastructure).

Decision:

  • Self-hosted: Llama 3 (70B) on your own servers.
  • Vector DB: Chroma (local).
  • Cost: $1,500/month (the GPU) + $200/month (hosting) = $1.7K/month.

Accepted trade-off: A high fixed cost (but guaranteed privacy).


Why this matters for an AI Engineer

1. Informed decisions

The Product Manager: "Why don't we use GPT-4 for everything?"

An AI Engineer (without understanding the trade-offs): "Okay, we'll use GPT-4."

An AI Engineer (understanding the trade-offs): "GPT-4 costs $60/10K requests. GPT-3.5 costs $2/10K. For a simple FAQ, GPT-3.5 is enough. Saving: 30×. I recommend a multi-model strategy."


2. Continuous optimization

Monitoring: "80% of the cost goes to GPT-4 on simple questions."

Action: Migrate the simple questions to GPT-3.5 → a 70% saving.


Summary

3 dimensions:

  1. Cost: GPT-4 (expensive) vs GPT-3.5 (30× cheaper).
  2. Latency: A simple chatbot (<1s) vs RAG (2-3s) vs Agents (5-10s).
  3. Quality: GPT-4 (high) vs GPT-3.5 (enough for simple tasks).

The trilemma: You can't maximize all 3 simultaneously.

Optimization strategies:

  1. A multi-model strategy (GPT-3.5 for simple, GPT-4 for complex).
  2. Intelligent caching (frequent FAQs → Redis).
  3. Prompt optimization (concise prompts).
  4. Batch processing (multiple requests in 1 call).

Decisions based on:

  • Budget (how much can you spend?).
  • Volume (1K or 1M requests/month?).
  • Required accuracy (90% or 95%?).
  • Tolerated latency (<1s or is 5s fine?).

Next step: Lesson 05: A Case Study — A complete design of a Q&A system (step by step).