In generative AI and prompt engineering, Few-shot, System, and Instruction refer to different ways of guiding or controlling a model’s behavior. Here’s what each term means:

Few-shot

  • Definition: A prompt that includes a few examples to help the model understand the task.
  • Usage: You provide 1–5 input-output pairs (shots) before the actual input.
  • Purpose: Helps the model generalize from examples.

Example:

Translate English to French:
English: Hello
French: Bonjour

English: Thank you
French: Merci

English: Good night
French:

System

  • Definition: A message or prompt that sets the global behavior or persona of the model.
  • Usage: Usually used at the beginning of a chat session to define the assistant’s role or style.
  • Purpose: Establishes tone, domain, constraints, etc.

Example:

You are a helpful assistant that explains complex legal concepts in simple terms.

Instruction

  • Definition: A direct command or instruction to the model within the user prompt.
  • Usage: Often used in zero-shot or one-shot formats, without extensive context.
  • Purpose: Tells the model exactly what to do.

Example:

Summarize the following paragraph in two sentences.

Summary:

Term Purpose Style Example Use
Few-shot Teach by example Sample-driven “Translate: Hello → Bonjour…”
System Set behavior/policies globally Meta-message “You are a math tutor.”
Instruction Command a specific action or task Direct prompt “Summarize this text…”

In Retrieval-Augmented Generation (RAG) scenarios, Query Rewriting and Query Summarization refer to techniques used to improve the effectiveness of information retrieval before generation. Here’s what they mean:

🔁 Query Rewriting

  • Definition: Modifying the original user query to make it more effective for retrieval.
  • Goal: Make ambiguous, incomplete, or conversational queries more specific and search-friendly.

Example: User query:

“What about Tesla in 2023?”

Rewritten query:

“What were the major developments related to Tesla Inc. in 2023?”

Why: Retrieval systems work better with precise, well-formed queries.

📝 Query Summarization

  • Definition: Condensing a long, complex query or multi-turn dialogue into a concise summary query.
  • Goal: Focus the retrieval on the core question or topic, especially in chat-like interactions.

Example: Conversation:

  • User: “Earlier you said Tesla had major events in 2023. Can you remind me what they were?”
  • Assistant: [retrieves and responds]

Summarized query:

“List of major events related to Tesla in 2023.”

Why: Helps retrieve the most relevant chunks without all the conversational baggage.

🔍 Why They Matter in RAG

RAG systems rely on retrieving relevant documents before generating an answer. Poor queries → poor retrieval → weak answers. So:

  • Rewriting improves clarity and specificity
  • Summarization improves focus and relevance

Both are forms of prompt engineering for retrieval, improving the model’s grounding in accurate information.

In Retrieval-Augmented Generation (RAG), Pre-generated answers refer to answers that are generated and stored in advance, rather than generated dynamically at runtime. Here’s what that means in detail:

🔹 What Are Pre-generated Answers?

They are precomputed responses associated with common or expected queries. Instead of generating an answer on the fly using retrieved documents, the system retrieves a ready-made answer.

🔹 Why Use Them?

  • ⚡ Speed: No need for generation; just retrieve and return.
  • 🎯 Consistency: Responses are curated and controlled.
  • 💾 Efficiency: Reduced computation load during runtime.

🔹 Where They Fit in RAG

In a typical RAG pipeline:

  1. User submits a query.
  2. The system retrieves relevant documents.
  3. A generation model creates a response based on those docs.

With pre-generated answers:

  • Step 3 is replaced by fetching a prewritten answer tied to the retrieved context or query.

🔹 Use Cases

  • FAQs
  • Customer support bots
  • Knowledge base with stable answers
  • High-traffic or repetitive queries

🔹 Example

Query Response Type
“What is your refund policy?” Pre-generated answer: “Our refund policy allows returns within 30 days…”
“Explain quantum mechanics” Generated dynamically from retrieved sources

So in summary: Pre-generated answers in RAG are optimized, stored responses for known queries—used to boost speed, accuracy, and scalability.