05. Zero-Shot, One-Shot & Few-Shot Prompting

flowchart LR
    subgraph ZERO["Zero-Shot"]
        Z1["Instruction only"] --> Z2["❌ Model guesses the format"]
    end

    subgraph ONE["One-Shot"]
        O1["Instruction + 1 example"] --> O2["✅ Follows the pattern"]
    end

    subgraph FEW["Few-Shot"]
        F1["Instruction + 3-5 examples"] --> F2["✅✅ Very accurate"]
    end

    ZERO -.-> ONE -.-> FEW

    style ZERO fill:#f59e0b,color:#fff
    style ONE fill:#3b82f6,color:#fff
    style FEW fill:#22c55e,color:#fff

✅ Good Zero-Shot:
"Translate the following English text to French:
'Hello, how are you?'"

✅ Good Zero-Shot:
"Classify this sentiment as positive, negative, or neutral:
'The product arrived broken and customer service was unhelpful.'"

flowchart TD
    subgraph ZERO_SHOT["Zero-Shot Performance"]
        COMMON["Common Tasks\nTranslation, Summarization\nClassification"] --> HIGH["✅ High accuracy"]
        NOVEL["Novel Tasks\nCustom formats\nSpecific schemas"] --> LOW["❌ Low accuracy"]
        AMBIGUOUS["Ambiguous Tasks\nVague instructions"] --> LOW2["❌ Unpredictable"]
    end

    style COMMON fill:#22c55e,color:#fff
    style NOVEL fill:#ef4444,color:#fff
    style AMBIGUOUS fill:#f59e0b,color:#fff

Task: Classify emails as 'Meeting', 'Task', 'Update', or 'Other'

Example:
Email: "Team standup at 10am tomorrow"
Classification: Meeting

Now classify:
Email: "The deployment is scheduled for Friday at 2pm"
Classification:

sequenceDiagram
    participant P as Prompt
    participant M as Model

    P->>M: "Classify emails. Example: X → Y"
    M->>M: "I see the pattern: email content → category label"
    M->>M: "The format is: 'Email: text' → 'Classification: label'"
    P->>M: "New email to classify"
    M->>P: "Classification: [label following the pattern]"

Classify the intent of each customer message.

Message: "I want to cancel my subscription"
Intent: Cancellation

Message: "Where is my order?"
Intent: Tracking

Message: "I was charged twice for the same order"
Intent: Billing Issue

Message: "Your product is amazing!"
Intent: Positive Feedback

Now classify:
Message: "I need to update my shipping address"
Intent:

flowchart TD
    subgraph TRADEOFF["Accuracy vs Token Cost Trade-off"]
        Z["Zero-Shot\nLowest Cost\nBaseline Accuracy"] --> O["One-Shot\nLow Cost\nGood Accuracy"]
        O --> F["Few-Shot\nHigher Cost\nBest Accuracy"]
        F --> D["Diminishing Returns\n10+ examples"]
    end

    style Z fill:#f59e0b,color:#fff
    style O fill:#3b82f6,color:#fff
    style F fill:#22c55e,color:#fff
    style D fill:#ef4444,color:#fff

flowchart TD
    Q1["Is this a common task?"] -->|Yes| Z["Use Zero-Shot\nSaves tokens"]
    Q1 -->|No| Q2["Can one example clarify?"]
    Q2 -->|Yes| O["Use One-Shot\nShows the pattern"]
    Q2 -->|No| Q3["Are there edge cases?"]
    Q3 -->|Yes| F["Use Few-Shot\n3-5 examples"]
    Q3 -->|No| O2["Use One-Shot\nSufficient for simple patterns"]

flowchart LR
    subgraph DYNAMIC["Dynamic Few-Shot"]
        INPUT["User Input"] --> EMBED["Convert to embedding"]
        EMBED --> SEARCH["Find similar examples\nfrom database"]
        SEARCH --> SELECT["Select top 3-5"]
        SELECT --> PROMPT["Build prompt with\ndynamic examples"]
        PROMPT --> LLM["LLM"]
        LLM --> OUTPUT["Response"]
    end

    style DYNAMIC fill:#8b5cf6,color:#fff

Zero-Shot: "Extract the date, amount, and vendor from this invoice"
→ Model might guess the format, get it partially wrong

Few-Shot:
Extract entities from invoice text.

Invoice: "Invoice #1234 from Acme Corp dated Jan 15, 2024 for $1,500"
Entities:
  - vendor: Acme Corp
  - date: 2024-01-15
  - amount: 1500
  - currency: USD

Invoice: "Receipt from WeWork - Monthly membership $299 - Feb 1 2024"
Entities:
  - vendor: WeWork
  - date: 2024-02-01
  - amount: 299
  - currency: USD

Now extract:
Invoice: "Payment of $89.99 to Netflix on March 5, 2024"
Entities:
→ Much more accurate

Zero-Shot: "Write a function to fetch data"
→ Could be any style, any language, any error handling

Few-Shot:
Generate TypeScript functions with error handling and JSDoc.

/**
 * Fetches a user by their ID
 * @param userId - The user's unique identifier
 * @returns The user object or null if not found
 */
async function getUser(userId: string): Promise<User | null> {
  try {
    const response = await fetch(`/api/users/${userId}`);
    if (!response.ok) return null;
    return await response.json();
  } catch (error) {
    console.error('Failed to fetch user:', error);
    return null;
  }
}

Now generate a function to fetch a product by SKU:
→ Follows the same pattern

Example 1: "What is Python?" → [result summary] [source]
Example 2: "How does DNS work?" → [result summary] [source]
Now: [user question] → [result summary] [source]