RAG, data quality, and professional prompt management

Section 5: RAG, data quality, and professional prompt management

Learning objectives

• Understand how RAG changes what the model actually sees
• Reduce hallucinations with constraints and sources
• Handle contradictions and multiple sources with clear priority
• Use meta-prompting, compression, and structured testing

RAG from a prompting perspective

RAG stands for Retrieval-Augmented Generation. Retrieval means look-up: rather than sending the whole knowledge base to the model, the system searches out the passages that appear to answer your particular question. Such a passage is called a chunk — a paragraph or section of a document, small enough to fit in the context window but large enough to stand on its own.

The chain, from uploaded document to answer:

1. Documents are uploaded to the knowledge base.
2. Chunking — the document is split into chunks. Each is tagged with where it came from: document, page, section.
3. Embeddings — each chunk becomes a numeric representation of what the text means. That is why a question about “holiday” can hit a passage that only mentions “leave”.
4. Vector index — those representations are stored in a searchable register.
5. Retrieval — your question is converted the same way, and the system fetches the chunks closest to it in meaning.
6. The selected chunks go into the context window along with your question.
7. The language model writes the answer from those passages.

What that means when you write the prompt: the model usually sees not the entire library, only the excerpts step 5 selected. Therefore:

• Clear questions → better matching to the right chunk. Vague questions retrieve vague passages.
• Use the same words and terms that appear in the material where you can — it raises the chance the right chunk is retrieved.
• Remind it of the big picture when needed: “summarise the document’s overarching conclusions” assumes retrieval actually picks representative parts — otherwise the summary risks being skewed.

Reduce hallucinations

Three approaches that tend to work:

1. Restrict the assistant to the supplied material where appropriate — and give it a ready-made phrase to use when the answer is not there. For example: “Answer only from the attached documents. If the answer is missing, write: That is not covered by the material.”
2. Require a source reference for claims that can be checked against the material, so you can go back and verify them.
3. Ask the assistant to say when it is unsure, rather than guess — especially when the material is thin or ambiguous.

Remember: models generate text — they don’t “know” whether they’re correct without grounding in something outside themselves.

Contradictions between sources

Ask the model not to pretend if the conflict is unknown. Example:

If sources give conflicting information: describe the conflict, name which documents are involved
and do not take a position on which version applies unless I have stated it.

You can also set priority: newer policy over older, governing documents over internal memos, etc.

Meta-prompting and compression

Meta-prompting: let the model first review your instruction (“what is unclear?”), suggest a v2, or generate a system prompt from a product brief. Always review the machine’s suggestion.

Compression: cut the politeness and framing that doesn’t steer the answer. “Could you please be so kind as to help me with…” gives the model nothing to act on — write the task and the format directly. The space in the context window does more good as instruction.

Output anchoring: give a filled-in template with headings for the model to complete – reduces free-form responses.

Tip: in Intric’s library you’ll find, among other things, a prompt expert you can import and iterate with.

Professional iteration and data types

• Test-driven prompting: define 3–5 typical success cases, a few ambiguous ones, and some out of scope – run after every change.
• Self-review: ask the model to rate and give feedback on a response according to criteria you care about.

Tables and exports: describe column meanings, date formats, and what empty cells mean before asking for analysis. Remove irrelevant columns – same principle as context engineering: less noise.

Intric has tools that help models with certain file types – but a clear prompt is still cheap insurance.

Summary

• RAG = retrieval of chunks, not magically reading everything.
• Hallucinations are countered with constraints, sources, and explicit uncertainty.
• Contradictions require policy and transparency.
• Meta-prompting, compression, and templates improve quality faster than just writing longer and longer prompts.

Congratulations – you’ve completed the advanced course. Keep testing in real cases and return to the modules when you introduce new teams or models.

Test your knowledge

1Question 1 of 42Question 2 of 43Question 3 of 44Question 4 of 4

Question 1 of 4

Question 1 of 4

What is a 'chunk' in a RAG pipeline?

What is a 'chunk' in a RAG pipeline?

• A summary of the entire document that is always sent in
• A piece of text that the system retrieves and sends to the model based on relevance
• A special XML tag for documents
• A synonym for hallucination

Back Next