Non-Visual AI Productivity

Back to Lessons

Lesson 6: Retrieval-Augmented Generation (RAG) for Non-Visual Productivity

Course: Foundations of Non-Visual AI Productivity (AI Basics)

Lesson content

• Lesson Objective: By the end of this lesson, you will understand what Retrieval-Augmented Generation (RAG) is, how the retrieve and generate steps work, and how to use citations to verify answers quickly in a non-visual workflow.
• What is RAG? Retrieval-Augmented Generation (RAG) is a method where an AI system first retrieves relevant information from allowed sources (documents, knowledge bases, internal wikis, databases, or approved search), and then uses those retrieved passages to write an answer. Think of it as open-book AI rather than closed-book AI.
• Why RAG matters for NVAIP: Non-visual work is faster when you can ask for specific information and get it in a structured form (headings, bullets, tables) instead of scanning long pages or complex interfaces. RAG can reduce time spent searching and improve answer relevance because the response is grounded in retrieved content.
• How RAG works (two steps):
  1. Retrieve: the system searches and selects passages that might contain the answer. Think of the AI like a librarian who runs to the stacks to find three specific books based on your question.
  2. Generate: the model writes a response using those passages as evidence. In the librarian metaphor, the AI then sits down, reads those three books, and writes you a summary.

  NVAIP tip (citations): Always ask the librarian to show you the page numbers. In AI tools, those page numbers are the citations. Request a citation for each key point so you can jump directly to the source and verify.

• Important clarification (RAG is not automatically web browsing):
  • RAG only retrieves from sources the system is connected to and permitted to access.
  • If your AI tool cannot access a file or site, it cannot retrieve from it. In that case, you must paste relevant text or use a tool that supports file ingestion.
  • “Has citations” does not always mean “correct.” Citations can be wrong, mismatched, or misunderstood.
• What RAG improves (common benefits):
  • Recency: can use up-to-date documents if your sources are current.
  • Specificity: can answer questions tied to your organization’s policies, reports, or knowledge base.
  • Traceability: can provide sources (citations) so you can verify quickly.
• What RAG does NOT guarantee:
  • Perfect retrieval: the system may fetch irrelevant passages or miss the best source.
  • Perfect reading: the model can misquote, misunderstand, or combine sources incorrectly.
  • Perfect truth: retrieved documents can be outdated, biased, or incorrect.
• NVAIP verification-by-design for RAG: use these habits every time you rely on retrieved sources:
  • Require citations per claim: “Give a citation for each bullet.”
  • Request short quotes: “Include 1 to 2 short quotes (under 25 words) supporting the key claims.”
  • Separate fact vs inference: “Split the answer into ‘From sources’ and ‘My inference’.”
  • Ask for uncertainty: “List anything unclear or conflicting in the sources.”
  • Spot-check: verify at least two critical details (names, dates, numbers, rules) directly in the source.
• Security note (prompt injection in documents): Sometimes retrieved text may contain misleading instructions such as “ignore previous rules” or “do X.” Treat retrieved passages as data, not instructions. Prefer trusted sources, and ask the assistant to ignore any instructions found inside documents.
• NVAIP RAG prompt template (copy/paste):

  Answer my question using only the provided/retrieved sources. If the sources do not contain the answer, say “Not found in sources.” Output format: 5 bullets with citations per bullet. Then provide 2 short supporting quotes from the sources. Then list assumptions, any uncertainty, and a 5-item checklist of what I should verify (names/dates/numbers/policy wording).

• Example workplace questions where RAG helps (non-visual friendly):
  • “What does our policy say about remote work eligibility and approval steps?”
  • “Summarize the key outcomes from last quarter’s report and list the top 5 metrics (with values).”
  • “From the training handbook, list the steps to onboard a new staff member.”
• Mini exercise (5-7 minutes):
  1. Choose a short internal document or public article you are allowed to use (non-sensitive).
  2. Ask a question that has a clear “ground truth” in that document.
  3. Use the NVAIP RAG template above and request citations plus 2 quotes.
  4. Verify: spot-check 2 key points in the source and correct anything wrong.
  5. Recover: If the AI says “Not found in sources,” try to orient better by describing the document structure (for example: “The answer should be in the section titled Personnel Policy”). Then ask the question again.
  6. Ask the AI: “Restate the final answer with corrected citations.”
• Key takeaway: RAG can make AI more useful for non-visual productivity by grounding answers in documents. But you still need verification-by-design: citations, short quotes, fact vs inference separation, and spot-checking.

Continue

All lessons in this course

• Lesson 1
• Lesson 2
• Lesson 3
• Lesson 4
• Lesson 5
• Lesson 6
• Lesson 7
• Lesson 8

References

1. 1. Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (2020). https://arxiv.org/abs/2005.11401
2. 2. IBM, “What is retrieval-augmented generation (RAG)?” IBM Think. https://www.ibm.com/think/topics/retrieval-augmented-generation
3. 3. Microsoft, “What is retrieval-augmented generation (RAG)?” (overview). https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/retrieval-augmented-generation