Skip to main content

Retrieval-augmented Prompts

Retrieval-augmented prompting integrates external search or database results into the prompt, allowing the language model to access up-to-date or domain-specific information beyond its training data. This approach is essential for tasks that require factual accuracy, current events, or specialized knowledge.

Key Concepts​

  • External Knowledge Integration: Incorporating retrieved documents, search results, or database entries into the prompt.
  • Context Expansion: Supplementing the model’s knowledge with relevant, up-to-date information.
  • Fact-based Reasoning: Enabling the model to ground its responses in external sources.
  • Dynamic Information: Accessing information that changes over time or is outside the model’s training window.

Best Practices​

  1. Curate Relevant Context

    • Select the most pertinent search results or documents.
    • Remove irrelevant or redundant information.

  2. Clearly Separate Sources

    • Use formatting or section headers to distinguish retrieved content from instructions or questions.
    • Cite sources or provide links when possible.

  3. Summarize or Highlight Key Points

    • Condense long documents to the most relevant excerpts.
    • Highlight facts or data points needed for the task.

  4. Prompt for Source-based Reasoning

    • Instruct the model to base its answer only on the provided context.
    • Ask for citations or references in the response.

Examples​

Basic Retrieval-augmented Prompt​

Context: "The Eiffel Tower is located in Paris and was completed in 1889."

Question: "When was the Eiffel Tower built and where is it located?"

Search Results Example​

Search results:
1. "OpenAI was founded in December 2015 by Elon Musk, Sam Altman, and others."
2. "OpenAI's mission is to ensure that artificial general intelligence benefits all of humanity."

Question: "Who founded OpenAI and what is its mission?"

Database Entry Example​

Database record:
Product: "SuperWidget 3000"
Release date: "2022-09-15"
Features: "Wireless, waterproof, 10-hour battery life"

Question: "What are the main features of the SuperWidget 3000?"

Multi-source Example​

Document 1: "Vitamin D is important for bone health."
Document 2: "Sunlight exposure helps the body produce vitamin D."

Question: "How can people get vitamin D and why is it important?"

Common Pitfalls​

  1. Irrelevant or Noisy Context

    • Including too many or unrelated search results.
    • Overwhelming the model with unnecessary information.

  2. Outdated or Incorrect Sources

    • Using stale or inaccurate data.
    • Failing to verify the reliability of sources.

  3. Ambiguous References

    • Not clarifying which source to use for which part of the answer.
    • Mixing information from conflicting sources without resolution.

  4. Ignoring Provided Context

    • The model answers from its own knowledge instead of the retrieved data.
    • Not grounding responses in the supplied information.

Use Cases​

  1. Question Answering

    • Factual queries
    • Current events
    • Domain-specific knowledge

  2. Summarization

    • Condensing retrieved documents
    • Multi-source synthesis

  3. Fact-checking

    • Verifying claims against external sources
    • Citing evidence

  4. Personalization

    • Using user-specific data from databases
    • Context-aware recommendations

When to Use Retrieval-augmented Prompts​

Retrieval-augmented prompting is ideal when:

  • The task requires up-to-date or domain-specific information.
  • Factual accuracy and source grounding are important.
  • The model’s training data is insufficient or outdated.
  • Multi-source synthesis or citation is needed.

When to Consider Alternatives​

Consider other techniques when:

  • The task can be completed with the model’s internal knowledge.
  • Reliable external sources are unavailable.
  • Simpler, standalone prompts are sufficient.

Tips for Optimization​

  1. Source Selection

    • Prioritize high-quality, relevant sources.
    • Remove duplicates and irrelevant entries.

  2. Context Formatting

    • Use clear labels and structure for each source.
    • Separate context from instructions and questions.

  3. Grounding Instructions

    • Direct the model to use only the provided context.
    • Ask for citations or evidence in the response.

  4. Iterative Testing

    • Experiment with different context lengths and formats.
    • Refine prompts based on model performance and accuracy.