Prompt Template

The static instructions are the fixed, non-changing text that defines the task, role, constraints, and desired output format for the LLM. This is the core logic of the template.

• Role Definition: You are a helpful customer support agent.
• Task Description: Summarize the following customer query into a single sentence.
• Output Formatting: Return your answer as a JSON object with keys 'summary' and 'sentiment'.
• Guardrails & Constraints: Do not use markdown. Do not make up information not present in the query.

Variables (or placeholders) are dynamic slots within the template that are populated at runtime with specific values. They enable a single template to handle infinite unique inputs.

• Syntax: Often denoted with curly braces, e.g., {user_query}, {{product_name}}.
• Purpose: Separate logic from data, allowing the same instruction set to process different documents, user inputs, or retrieved contexts.
• Example: In a summarization template, {document_text} is replaced with the actual article to summarize.

Structured formatting uses clear visual separators to organize different sections of the prompt, helping the LLM parse instructions from data and examples.

• Common Delimiters: Triple backticks (```), XML tags (<context>...</context>), section headers (## EXAMPLES ##), or horizontal rules (---).
• Function: Prevents instruction bleeding, where the model confuses examples or user input as part of the core command.
• Best Practice: Consistently using delimiters improves the model's ability to follow complex, multi-part templates reliably.

Few-shot examples are input-output pairs embedded within the template to demonstrate the desired task to the LLM via in-context learning. They are a powerful method for steering model behavior without fine-tuning.

• Structure: Each example typically shows a User: input and the expected Assistant: output.
• Impact: Provides a pattern for the model to follow, dramatically improving performance on tasks requiring specific formatting or reasoning.
• Consideration: Examples consume context window tokens, so they must be chosen carefully for maximum instructional value.

Inference parameters are model configuration settings often specified alongside or within a prompt template to control the generation process. While not part of the prompt text itself, they are critical for deterministic output.

• Key Parameters:
  • Temperature: Controls randomness (e.g., temperature=0.1 for focused, reproducible answers).
  • Top-p (Nucleus Sampling): Limits token selection to a probability mass (e.g., top_p=0.9).
  • Max Tokens: Sets a hard limit on response length.
• Template Integration: These parameters are typically managed in the application code that renders the template, ensuring consistent generation quality.

The context integration slot is a specialized variable designed to hold dynamically retrieved information, forming the core of a Retrieval-Augmented Generation (RAG) architecture. It grounds the LLM's response in external, verifiable data.

• Purpose: To inject relevant facts, documents, or database records into the prompt, reducing hallucinations.
• Template Example: Using only the following context: {retrieved_context} Answer the question: {user_question}
• Engineering Consideration: The slot's placement and the instructions around it are critical for ensuring the model prioritizes the provided context over its internal knowledge.

This template provides a clear, direct command to the model without examples. It relies on the model's pre-trained knowledge and is ideal for simple, well-defined tasks.

• Structure: A single, concise instruction.
• Use Case: Classification, summarization, or simple transformation tasks.
• Example: Translate the following technical document from English to French: {document_text}

This template includes several input-output examples within the prompt to demonstrate the desired task format and logic before the model processes the new input.

• Structure: Instruction, followed by 2-5 labeled examples (Input:, Output:), then the target input.
• Use Case: Teaching complex formatting, specific stylistic tones, or nuanced classification.
• Example: For sentiment labeling: Classify the sentiment of these reviews. Input: "The battery life is exceptional." Output: Positive Input: "The interface feels clunky." Output: Negative Input: {new_review} Output:

This template defines a specific persona, expertise level, and operational context for the model, framing its entire response strategy.

• Structure: A system-level role definition, optional context/constraints, then the user query.
• Use Case: Customer support agents, domain-specific analysts (e.g., legal, medical), or creative writing.
• Example: You are an experienced cybersecurity analyst. Explain the concept of a zero-day exploit to a non-technical executive. Use simple analogies and avoid jargon.

This template explicitly dictates the format of the model's response, often using JSON, XML, or markdown schemas, ensuring machine-parsable outputs.

• Structure: Instruction, a detailed schema definition with placeholders, then the data to process.
• Use Case: Data extraction, API response generation, populating databases.
• Example: Extract the company name, date, and monetary value from the following contract clause. Return a valid JSON object with the keys: "company", "date", "amount". Clause: {clause_text}

This template instructs the model to articulate its intermediate reasoning steps before stating a final answer, significantly improving performance on complex logic, math, or planning tasks.

• Structure: A problem statement followed by an instruction like "Let's think step by step" or inclusion of few-shot examples showing the reasoning process.
• Use Case: Mathematical word problems, multi-step planning, causal reasoning.
• Example: A bakery sells cookies in boxes of 6. If they have 17 boxes, how many cookies do they have? Let's think through the reasoning step by step.

This template dynamically incorporates retrieved information from an external knowledge base (like a vector database) into the prompt to ground the model's response in factual, up-to-date context.

• Structure: Instruction, a Context: section containing retrieved documents, then the query that must be answered based solely on that context.
• Use Case: Question answering over proprietary documents, reducing hallucinations, providing citations.
• Example: Answer the question based only on the provided context. Context: {document_1} {document_2} Question: {user_question}