How RAG turns "I'm not sure" into a cited answer

A real scenario — an engineer asking why the office network is slow — walked through scene by scene.

1

The question

An engineer asks the company's internal AI assistant a question that comes up in almost every interview about "what could go wrong in production": the network is misbehaving.

"Video calls keep freezing on the 3rd floor around 2pm every day. What should I check first?"
2

Without RAG — generic, and it shows

Ask a plain LLM with no access to your company's actual documents, and it can only draw on generic knowledge from training — it has never seen your office's floor plan, your router model, or last month's ticket history.

"This is usually caused by network congestion or Wi-Fi interference. Try restarting your router, moving closer to the access point, or checking for background downloads."

Not wrong, exactly — just not useful. It's the same answer it would give any company, in any building, on any day.

WITH RAG
3

Retrieval — search the real handbook first

Before answering, the system searches the company's own "Network Troubleshooting Handbook" — a real internal document — for passages related to the question, and pulls back the most relevant one.

network-troubleshooting-handbook.md — §3.2
Bandwidth contention: the 3rd floor shares an uplink with the render farm. Between 13:30–15:00, nightly render jobs are re-queued during business hours, saturating uplink bandwidth and degrading real-time video/voice traffic. Confirm via switch port counters before assuming a Wi-Fi issue.
4

Assembly — question + retrieved passage, together

The retrieved passage isn't shown to the user directly — it's quietly added to the prompt alongside the original question, so the model answers with real context instead of guessing.

Prompt sent to the model: "[handbook §3.2 passage above] — Using only the passage above, answer: Video calls keep freezing on the 3rd floor around 2pm every day. What should I check first?"
5

Grounded answer, with a citation

Same question, same model — but now it can point at the actual cause and where it came from, not a guess.

"Check switch port counters on the 3rd-floor uplink around 13:30–15:00 first — the render farm's nightly jobs get re-queued into business hours and saturate that link, which is the likely cause of the video freezing. This is a known pattern, not a Wi-Fi problem."
📄 Source: network-troubleshooting-handbook.md, §3.2

That mechanism is called RAG

Retrieval-Augmented Generation: search your own documents first, then let the model answer using what it found — instead of hoping it already knows. This site runs a real, working implementation of it.

See the real local RAG app →