LAUNCHMongoDB 8.3 is built for the sub-100ms retrieval & zero downtime AI demands. Read blog >
AI DATAStop fighting your data layer. Get the memory & retrieval agents need to scale. Read blog >

What is Prompt Engineering?

Imagine you told your friend you needed an orange and she gave you a grape. Unfortunately, this happens all the time with large language models (LLMs) when prompts aren’t optimized. In fact, ineffective prompts can greatly degrade the outcomes you receive from generative AI systems. Prompt engineering is the practice of ensuring that when you ask your friendly generative AI chatbot for an orange that you get an orange.

The evolution of artificial intelligence (AI) over the past few decades has been immense, evolving from rudimentary computing systems to the sophisticated LLM-powered, generative AI models of today. At the core of AI and its potential is a deceptively complex, yet powerful and relatively new discipline itself: prompt engineering. But what was a niche area of exploration for AI researchers and practitioners is becoming an increasingly essential tool for non-technical professionals using the current generation of generative AI tools for a wide range of business tasks.

Table of contents

A short history of AI and prompt engineering

The story of AI begins in the 1950s, with the development of early computing systems and the foundational work of pioneers like Alan Turing. The Turing Test, proposed in 1950, set the stage for what would become the quest for true artificial intelligence—machines that could mimic human reasoning. Fast forward to the 1980s and 1990s, when the field of machine learning began to take shape, leading to more dynamic AI models. The 1997 victory of IBM's Deep Blue over chess grandmaster Garry Kasparov was a symbolic milestone, showcasing AI's potential in more complex problems and problem-solving. The early 2000s saw the rise of generative AI, marked by the development of algorithms capable of creating content—from text to images. This era witnessed significant advancements in natural language processing (NLP), laying the groundwork for more sophisticated language models.

The 2010s brought about a revolution with the introduction of large language models like GPT (Generative Pre-trained Transformer) and BERT (Bidirectional Encoder Representations from Transformers). These AI models, trained on vast amounts of data—and popularized by OpenAI's ChatGPT—are now used to generate coherent and contextually relevant text, pushing the boundaries of what AI could achieve.

As AI models have become more mainstream and accessible, organizations are shifting focus and resources from developing proprietary models to ensuring interactions with generative AI yield the desired results. As a result, prompt engineering has become increasingly crucial, emerging not just as a skill but as a vital discipline in our current AI era.

From basic syntax to NLP

Though a relatively new term, the roots of prompt engineering begin with the early developments of artificial intelligence and natural language processing. The story of prompt engineering starts with the birth of computational linguistics in the 1950s. Initially, the focus was on programming computers to understand basic language syntax and semantics. However, the interaction with these early systems was limited and rudimentary.

In the 1960s and 70s, researchers began experimenting with more advanced models of language understanding, leading to the development of ELIZA in 1966 by Joseph Weizenbaum. ELIZA was a breakthrough because it demonstrated a form of interaction where the computer's responses were contingent on the user's input, laying the groundwork for what would evolve into prompt engineering. The 1980s witnessed a surge in rule-based AI systems, where the effectiveness of the interaction was highly dependent on the precision of the input commands or prompts. This era was crucial in highlighting the relationship between the input (prompt) and the AI model's ability to output, setting the stage for more advanced prompt-based interactions. The 1990s saw significant advancements in natural language processing (or NLP). With the introduction of more nuanced AI models, the role of the prompt and prompt engineering began to gain prominence. AI systems could now handle more complex inputs, and the quality of these inputs started to directly influence the quality of the outputs.

Enter machine learning (ML), big data, and LLMs

The advent of machine learning and big data in the 2000s revolutionized AI, leading to more powerful and adaptable models. The introduction of large datasets and the ability of AI to learn from these datasets meant that prompts needed to be more refined and contextually aware. This period marked the shift from static, rule-based prompts to more dynamic and nuanced prompt engineering.

Today, we are entering a new generative AI era. The launch of large language models, like GPT and BERT in the 2010s, marked a new era in AI, where the complexity and capabilities of AI systems and AI language models reached unprecedented levels. Prompt engineering evolved into a critical discipline, with prompt engineers becoming key players in guiding AI responses to be more accurate, contextually relevant, and creative. Today, generative AI tools are becoming a commonplace "assistant" across many industries around the world.

The emergence of retrieval-augmented generation (RAG) and vector search technologies has added a powerful dimension to prompt engineering. RAG integrates retrieval mechanisms with generative models, enhancing the contextual relevance of AI outputs by leveraging vast databases of structured and unstructured data. MongoDB Atlas, with its robust capabilities in vector search, plays a pivotal role here. Vector search enables AI systems to find semantically similar documents, thus refining the context provided to the generative models. This technology moves beyond traditional keyword searches, utilizing high-dimensional vector spaces to understand the deeper relationships between data points.

Vector search’s ability to handle complex and varied data forms—ranging from text and images to audio—complements RAG models by providing enriched, context-aware information. This integration significantly improves the accuracy and relevance of AI-generated content, making prompt engineering even more critical. The AI systems can now generate more precise and contextually relevant responses, thanks to the sophisticated data retrieval methods.

What are some specific examples of prompt engineering use cases?

Use case: content creation

When it comes to digital media and journalism, prompt engineers are crafting AI inputs that generate creative and engaging articles, blog posts, and scripts. These AI-generated pieces can mimic specific writing styles or adhere to particular thematic guidelines, providing a new tool for content creators to expand their reach and efficiency.

Use case: customer service

The world of customer service has been particularly revolutionized by prompt engineering. By designing sophisticated prompts, engineers are enabling chatbots and virtual assistants to provide more accurate, contextually relevant, and empathetic responses to customers. AI not only enhances customer experience but also streamlines operations, allowing human agents to focus on more complex tasks and queries.

Use case: education and training

In education, prompt engineering is used to create personalized learning experiences. AI can generate custom-tailored study materials, quizzes, and interactive learning modules based on specific prompts. This personalization can accommodate different learning styles and paces, making education more accessible and effective.

Use case: healthcare

In healthcare, prompt engineering assists in creating AI tools that can interpret and respond to patient inquiries, aid in diagnostics, and even assist in mental health therapy sessions. By crafting prompts that are sensitive to medical contexts and patient needs, AI can become a valuable tool for healthcare professionals.

Use case: legal and compliance

The legal field benefits from prompt engineering through the creation of AI systems that can parse through vast amounts of documents to assist in research or identify compliance issues. This not only saves time but also enhances the accuracy and thoroughness of legal work.

Use case: marketing and advertising

In marketing, AI, guided by well-engineered prompts, is used for creating targeted advertising content, analyzing market trends, and generating customer insights. This allows for more personalized and effective marketing strategies.

Use case: entertainment

In the entertainment industry, prompt engineering is being used to create novel interactive experiences. AI can generate music, artwork, or even plot lines for games and movies based on specific prompts, opening new frontiers for creative expression.

Use case: language translation and localization

Prompt engineers are also contributing to the field of language translation, where AI is used to not just translate text but also localize content culturally and contextually, making global communication more seamless.

What are some prompt engineering best practices?

Understanding how AI models interpret and process language is key to success. The perfect prompt can create truly insightful, tailored answers rather than generic responses.

But prompt engineering is more than just asking questions. It's about understanding how AI models interpret and process language. The right prompt can mean the difference between a generic response and a thoughtful, personalized answer. Effective prompt engineers have a creative streak, LLM-specific technical knowledge, and an understanding of how models "think."

Tips for writing effective prompts

Be clear and specific: Vagueness can lead to irrelevant answers. Ensure your prompts are clear and direct.

  • Why specificity matters: Large language models are trained on massive amounts of text data. They can generate a wide range of responses, but they might struggle to pinpoint your exact needs if your prompts are too open-ended. Ambiguity increases the risk of misinterpretation and irrelevant output.
  • Pinpoint your intent: Ask yourself, "What specific information do I want to extract, or what precise action do I need the model to take?" Distill your goal into a focused request.

Example: vague vs. specific

  • Vague: "Tell me about customer data"
    • Problem: What aspect of customer data? Purchase history? Demographics? Support tickets? This leaves the model guessing.
  • Specific: "Summarize the average order value for customers located in California during the last quarter."
    • Improvement: A clear target (average order value) and constraints (location, time period) give the model a much better focus.

MongoDB-specific tips

  • Understand your schema: When interacting with MongoDB data, reference fields and collection names precisely.
  • Use filtering: Employ MongoDB query syntax within the prompts to focus the model's search: "Find customers with purchase totals greater than $500 from the 'orders' collection."
  • Avoid broad questions: Instead of "What insights can I get from my sales data?" frame it as, "Generate a list of the top 10 selling products by revenue for the previous month."

Context matters: Provide enough background information. Context helps the AI understand the nuances of the request.

  • RAG and the importance of context: In RAG models, the ability to retrieve relevant information from a knowledge base is key. Providing sufficient context helps the model narrow down its search to the most pertinent documents or passages.
  • Focus on pinpointing the knowledge source:
    • Internal knowledge base: If you're using a custom knowledge base, be specific about the document types or collections it contains.
    • External sources: If the model can access web data, guide it with keywords or even specific URLs if possible.
  • Context for different RAG tasks:
    • Question answering: Frame your question, then add background on the topic or relevant document content.
    • Summarization: Provide details about the source document and the desired focus of the summary.
    • Generative text (based on retrieved info): Supply a knowledge snippet and the type of text you wish to generate (e.g., a product description or an email draft).

Here’s an example to illustrate:

  • Task: Generate a product description for a new item in your MongoDB product catalog.
  • Poor context: "Write a product description."
  • Improved context:
    • "Product Name: [Name from MongoDB]. Product Category: [Category from MongoDB]. Target Audience: [Describe]. Here's a features list from my 'products' collection: [list key features]."

MongoDB-specific tips for RAG

Harness the power of your MongoDB data to get the most out of retrieval-augmented generation models: MongoDB’s compatibility with JSON documents offers a distinct advantage, enhancing the efficiency and accuracy of RAG implementations.

Here’s how you can make the most of your MongoDB data with RAG:

  • Structure your data for retrieval:
    • Relevant fields: Include fields the model will likely need for context and generation tasks. Consider key product attributes, customer demographics, or relevant text fields in your documents.
    • Descriptive document titles: Use clear, descriptive titles to improve document retrieval when RAG models search across your database.
  • Leverage MongoDB Atlas Search:
    • Rich indexing and search capabilities: Atlas Search offers robust indexing and search capabilities, making it an ideal companion for RAG models. This combination provides fast and accurate retrieval of relevant documents.
    • Embed MongoDB queries: Enhance prompts with embedded MongoDB queries for precise filtering and knowledge retrieval directly from your collections.
  • Utilize aggregation pipelines:
    • Pre-calculate results: If you need RAG-generated summaries of complex aggregations, pre-calculate and store results using MongoDB aggregation pipelines. This ensures easy access by the model and reduces on-the-fly computation.
  • Incorporate Vector Search:
    • Contextual and semantic searches: Vector search, a critical technology for RAG context, allows for searching based on the similarity of data rather than exact matches. MongoDB Atlas Vector Search is particularly suited for this, enabling more nuanced and accurate retrievals in RAG implementations.

Here’s an example using product recommendations with RAG + MongoDB Atlas Search.

  1. User search: The customer types in a partial product name.
  2. RAG with Atlas Search: The model retrieves potentially relevant product documents from MongoDB using Atlas Search.
  3. Context expansion: RAG accesses order history from the 'customers' collection to understand the user's past preferences.
  4. Recommendation generation: The model combines search results with purchase context, then generates descriptions of the most likely 'next best product' for the user.

Keep it structured: Structured prompts resemble the way humans communicate logically. They help the AI model break down information and process it in a step-by-step manner. Here are some ways to structure your prompts:

  • Bullet points or numbered lists: Break down complex requests into smaller, more manageable components. This is especially helpful for tasks with multiple requirements.
    • Example: To generate a marketing email, a structured prompt might look like:
      • Target audience: Small business owners interested in productivity tools.
      • Key benefits to highlight: Time-saving, ease of use, collaboration features.
    • Call to action: Start a free trial.
  • Clear questions: Explicit question formats help the model zero in on an answer. This aligns the response with your desired outcome.
    • Example: Instead of "Tell me about my website traffic," ask, "What were the top 3 sources of website traffic last week?"
  • Templates or frameworks: Using consistent skeletons for common tasks can improve prompt quality and speed up development.
    • Example (AIDA framework for ad copy):
      • Attention: [Grab the reader's attention with a compelling statement]
      • Interest: [Highlight unique benefits of your product]
      • Desire: [Spark the user's desire for your solution]
      • Action: [Include a clear call to action like "Visit our website"]

MongoDB-specific tips

Atlas Search provides powerful indexing and querying capabilities for your data within MongoDB. Harness these features to enhance your prompt engineering and get laser-focused results:

Iterative approach: Prompt engineering is rarely a one-and-done process. Start with an initial prompt and then iteratively refine it based on the AI's response and your goals. Consider these techniques:

  • Evaluate and refine: Analyze the model's output. Did it fully address your request? Was there a lack of specificity or a misunderstanding? Tweak your wording, focus, or structure accordingly.
  • Experiment and chain prompts: Explore different ways to phrase your request, or break complex tasks into a series of smaller, chained prompts for step-by-step instructions.
  • Human feedback loop: Incorporating a human-in-the-loop approach, where a person reviews and rates model responses, can help fine-tune prompts over time and improve the model's understanding of your requirements.

Example of iterative prompting:

Goal: Get a concise and engaging summary of a lengthy customer support ticket.

  • Initial prompt: Summarize this customer support ticket.
    • The model might provide a long summary that still needs editing.
  • Refinement 1: Summarize the main customer issue and resolution in this support ticket.
    • The model might get closer, but the tone could be too technical.
  • Refinement 2: Provide a customer-friendly summary of the issue and solution in this ticket. Limit it to 2-3 sentences.
    • The model likely produces a more suitable output for direct use.

Iterative approach with MongoDB Atlas Search

Since your success with prompt engineering and Atlas Search depends on the quality of your queries and settings, consider these iterative techniques specifically for search-driven applications:

  • Adjust search relevance scores: Analyze retrieved results and the scores assigned by Atlas Search. Is the model prioritizing the most relevant documents? Experiment with field weights or boosting techniques to fine-tune your results ranking and improve the quality of information supplied to the AI in subsequent steps.
  • Refine search operators: If the model isn't finding the exact results you were after, look at the operators in your embedded search queries. Are you using the most appropriate ones? Should you include more or less restrictive operators for better focus?
  • Analyze misinterpretations: If the model's output is misaligned with search results, there might be a mismatch between your prompt's wording and the way Atlas Search is indexing your data. Consider alternative phrasing or employ synonyms to improve alignment.
  • Iterate on search configuration: Experiment with different analyzers, stop words, and other Atlas Search settings to see how they influence the results delivered to the model.

Here’s an example to illustrate:

Let's say your goal is to generate customer support insights from your product feedback.

  1. Initial Atlas Search query: "Find feedback documents mentioning 'bug'"
  2. Analysis: Results might be too broad
  3. Iteration 1: "Find feedback documents mentioning 'bug' where product: 'Widget X'" (Add product specificity)
  4. Analysis: Better, but perhaps too many results from the past
  5. Iteration 2: "Find feedback documents mentioning 'bug' where product: 'Widget X' AND date > YYYY-MM-DD" (Time constraint)

Prompting models and how to use them

To unlock the full potential of AI language models, understanding and using prompting models is essential. Prompting models offer a structured and strategic approach to guiding these models, enabling us to extract the most valuable and insightful responses.

By understanding the nuances of these models, you can unlock tailored information, streamlined problem-solving, and audience-specific communication.

Prompt engineering practitioners often use one of the following prompt models:

  • Chain-of-thought prompting
  • Generated knowledge prompting
  • Complexity-based prompting

Chain-of-thought prompting

This technique involves providing intermediate reasoning steps for the model to follow.

How it works: Guides the AI through step-by-step reasoning, breaking down complex problems into smaller sub-tasks

Benefits:

  • Improves model accuracy on complex tasks
  • Provides transparency into the AI's decision-making process
  • Helps the model avoid logical fallacies

Best for:

  • Problem-solving scenarios where the solution path matters
  • Situations where you need the AI to justify its answers
  • Debugging the model's logic and identifying areas for improvement

Generated knowledge prompting

This approach focuses on leveraging the AI's knowledge base to perform certain tasks.

How it works: Leverages the AI's expansive knowledge base to extract information, generate summaries, or synthesize insights

Benefits:

  • Access to vast amounts of pre-processed information
  • Useful for rapidly obtaining factual summaries or overviews of topics
  • Can aid in the creation of original content based on existing knowledge

Best for:

  • Research and knowledge-gathering tasks
  • Summarizing complex topics for easy digestion
  • Serving as a starting point for further exploration or creative work

Complexity-based prompting

This technique emphasizes tailoring prompts to match the audience's expertise or the desired level of detail in the AI's response.

How it works: Adapts the prompts to match the audience's expertise or the desired level of detail in the output

Benefits:

  • Creates tailored communication, avoiding misunderstandings
  • Improves accessibility of complex concepts for broader audiences
  • Can provide multiple levels of explanation, from high-level to granular

Best for:

  • Scenarios involving diverse audiences with varying technical backgrounds
  • Educational or onboarding materials where the user's level of knowledge may be unknown
  • Customer support, where responses might range from simple to detailed, depending on the user's technical ability

Additional notes:

Hybrid approaches are common: Prompt engineers often blend these techniques to achieve optimal results. For example, a chain-of-thought prompt might include elements of generated knowledge prompting if the AI needs to access factual data at a certain step.

Consider the model's strengths: Some large language models might be better suited to specific prompting styles. Experimentation is key to finding the best model/prompt combinations.

Is there a demand for prompt engineers?

While it started out as a specialized role for data scientists, prompt engineering is no longer just for traditional tech roles. It may even soon be a career path.

As more and more people use generative AI, professionals from fields as diverse as writing, education, marketing, and even psychology will bring unique perspectives to AI interactions. The varied understanding of language nuances, audience engagement, and human behavior can significantly enhance the effectiveness of AI prompts.

More to the point, the demand for prompt engineers is not just hypothetical—it's a developing trend in the job market. According to a 2023 McKinsey Global Survey, 7% of respondents from organizations that have adopted AI reported actively hiring prompt engineers.1 This data point not only underscores the growing recognition of the role's importance but also suggests a burgeoning market for these skills​​.

Companies are increasingly aware that the quality of inputs into generative AI models significantly influences the outcomes. You want relevant responses. Consequently, they are offering competitive salaries for prompt engineers, recognizing the value they bring to optimizing AI interactions. This trend is a strong indicator of the evolving AI landscape, where the ability to effectively communicate with and through AI systems is becoming as valuable as the technology itself.

What is the future of AI and prompt engineering?

AI has arrived. Organizations around the world are rushing to stay ahead of the pack with the latest generative AI applications.

Prompt engineering is an important element of AI. And crafting an effective prompt is an art in itself. It involves clarity, specificity, and an understanding of the AI's capabilities. A good prompt should be clear enough to guide the AI but not so restrictive that it stifles creativity or flexibility. It should take into account the context in which the response will be used and be designed to elicit the most relevant, accurate information. Understanding the target audience, the intended use, and the limitations of the AI system are all crucial in crafting a prompt that yields useful and meaningful outputs.

As AI continues to evolve and become more ingrained in our daily lives, the ability to effectively communicate with these systems will become increasingly important. With its wide range of applications and the growing demand for skilled professionals, prompt engineering is not just a niche skill. It’s an essential expertise in our rapidly evolving AI era.

Prompt engineering and MongoDB

The rise of generative AI and powerful language models signals a watershed moment. Prompt engineering, once a niche discipline, is fast becoming a cornerstone skill for leveraging the true potential of these systems. Here are key trends shaping the future:

  • Prompting as programming: As AI models become more sophisticated, the prompts themselves will resemble a new kind of programming language—one focused on natural language communication of intent rather than rigid code structures.
  • Democratization of AI: Well-designed prompting tools and interfaces could lower the barrier to entry for AI applications. Non-technical users may be able to generate marketing copy, write simple scripts, or extract insights from data without needing traditional coding skills.
  • Human-AI collaboration: Prompt engineers are the bridge-builders between human needs and AI capabilities. They will play a vital role in refining and interpreting outputs, translating complex requests into AI-understandable formats and ensuring the ethical use of these powerful tools.
  • The rise of specializations: We'll likely see prompt engineers focusing on specific domains. For instance, experts in healthcare, law, or finance will fine-tune their prompting skills for the unique requirements of their industries.

The combination of prompt engineering and MongoDB offers exciting possibilities:

  • Unlocking data insights: Using natural language to query and summarize complex data stored in MongoDB could make analytics accessible to wider teams within an organization.
  • Seamless data-driven content: Generate reports, product descriptions, or support materials directly informed by your MongoDB data, ensuring content stays up-to-date and relevant.
  • Intelligent search interfaces: Build conversational search experiences on top of MongoDB Atlas Search, where users can ask questions in plain language and get precisely the information they need.
  • Advanced automation: Trigger MongoDB pipelines or workflows based on insights extracted from customer feedback, logs, and other data sources using nuanced prompts.

As AI continues to evolve, so will the art and science of prompt engineering. Those who master it will be at the forefront of shaping how we interact with intelligent systems, unlocking new possibilities and value everywhere AI is applied.

Learn more about MongoDB Atlas Search.

Get Started With MongoDB Atlas

Try Free