Lesson 43 of 1550
RAG For Ops Manuals: Retrieval That Actually Retrieves
Retrieval-Augmented Generation lets you ground answers in your own ops manuals. Most RAG systems fail not at generation but at retrieval — here's how to fix that.
Lesson map
What this lesson covers
Learning path
The main moves in order
- 1RAG is mostly retrieval, barely generation
- 2RAG
- 3chunking
- 4embedding
Concept cluster
Terms to connect while reading
Section 1
RAG is mostly retrieval, barely generation
When a RAG system gives a wrong answer, the LLM is rarely the culprit. The retrieval step pulled the wrong passages, and the LLM faithfully summarized them. Improving generation prompts won't fix this. Improving retrieval will.
The chunking question
- 1Chunk by section header, not by character count, when the doc has structure
- 2Add a parent-document reference to every chunk so the model can ask for more context
- 3Don't split tables across chunks
- 4For long procedures, keep numbered steps together
Reranking changes the game
Embedding similarity gets you 'topically close.' That's not the same as 'answers the question.' A reranker — even a small one — re-scores the top-50 retrieved chunks against the actual query. Adding a reranker is often a 20-30 point retrieval-quality jump.
Compare the options
| Symptom | Likely cause | Fix |
|---|---|---|
| Right topic, wrong specifics | Chunks too small, missing context | Bigger chunks or parent-doc lookup |
| Hallucinated steps | Retrieval missed the actual procedure | Reranker, better embeddings |
| Outdated answer | Stale chunks not re-indexed | Scheduled re-embedding job |
| Confidently wrong | Generation prompt not strict enough | Force 'answer only from passages' grounding |
The big idea: RAG quality is retrieval quality. Build the eval set, then tune retrieval, then worry about the LLM.
End-of-lesson quiz
Check what stuck
15 questions · Score saves to your progress.
Tutor
Curious about “RAG For Ops Manuals: Retrieval That Actually Retrieves”?
Ask anything about this lesson. I’ll answer using just what you’re reading — short, friendly, grounded.
Progress saved locally in this browser. Sign in to sync across devices.
Related lessons
Keep going
Adults & Professionals · 11 min
Internal Document RAG: Making the Wiki Actually Useful Again
Most company wikis are graveyards of stale info. AI RAG systems can resurrect them — when paired with content-freshness tracking and source citation.
Adults & Professionals · 40 min
SOP Automation: Turning Tribal Knowledge Into Prompted Workflows
Standard Operating Procedures live in PDFs nobody reads. An LLM can compile them into living, prompt-driven checklists that adapt to context.
Adults & Professionals · 10 min
Ticket Triage With LLMs: Routing Without The Backlog
Support and ops queues drown teams in repetitive sorting work. A well-prompted LLM classifier can do 80% of that triage with confidence-aware handoff.