RAG And Knowledge Systems

What To Know

RAG appears often because many customer environments have scattered operational knowledge: docs, tickets, Slack threads, setup notes, SOPs, logs, and tribal knowledge.

The goal is not magic chat. The goal is cited, inspectable help that reduces search time and makes workflows repeatable.

Core Concepts

• Source ingestion.
• Cleaning and normalization.
• Chunking.
• Embeddings.
• Vector search.
• Metadata filters.
• Reranking.
• Citations.
• Evaluation queries.
• Human feedback.

Common Failure Modes

• The right source was never ingested.
• Chunking split important context.
• Retrieval finds similar language but wrong procedure.
• The model answers beyond the sources.
• Citations are too vague.
• Old docs conflict with newer docs.
• Permissions are ignored.

Interview Angle

Strong phrasing:

For knowledge systems, I care about source quality, retrieval quality, citations, uncertainty, and whether the answer helps a real workflow.