What is context injection?
The controlled assembly of retrieved evidence into a structured, model-ready prompt context.
Interview Questions: Context Injection
Why is raw chunk stuffing risky?
It creates noise, citation confusion, token waste, and prompt-injection exposure.
What belongs in a context pack?
Question, evidence text, citation ids, metadata, answer rules, budget decisions, and dropped-context reasons.
How do you order evidence?
Prefer direct relevance, authority, freshness, diversity, and logical answer flow.
How do you defend against prompt injection?
Separate instructions from evidence and treat retrieved text as quoted data, never as system instruction.
What makes a good citation?
A stable source id tied to source, section, version, and chunk or quote span.
How do you handle context overflow?
Reserve answer budget, drop duplicates/low-authority chunks, log reasons, and refuse if evidence becomes insufficient.
What can compression break?
It can remove conditions, exceptions, dates, thresholds, or attribution needed for faithful answers.
What metrics matter?
Context precision, citation support, token usage, truncation rate, conflict rate, and answer groundedness.
What is the senior answer?
Context injection is where retrieval quality becomes answer quality through structured evidence, boundaries, citations, and observability.