Most screening agents are built as interviewers: a list of questions, asked in order. They work fine right up until the candidate asks something back — and candidates always ask something back.
"What's the salary range?" "Is this remote?" "Which team is it for?" "How many rounds are there?"
An agent with no answer does one of two bad things. It deflects everything to "a recruiter will follow up", which makes the call feel like a form. Or it invents an answer, which is considerably worse — a hallucinated salary range is a real commercial problem.
What a knowledge base does
It gives the agent a body of facts it is allowed to answer from: the job description, the interview process, the benefits, the office locations, the common objections and their honest answers.
The mechanism matters less than the constraint. The useful property is not "the agent knows more" — it is "the agent has a defined set of things it may state, and everything else is explicitly a follow-up". That boundary is what stops invention.
What to put in it
Write it as answers, not as documents. The failure mode is uploading a 30-page HR handbook and expecting the agent to find the relevant sentence mid-call.
What works is a plain list of the questions people actually ask, each with the answer you would want a recruiter to give:
- Compensation. The range you are willing to state, and what to say if they push beyond it.
- Location and remote policy. Specifically — "hybrid" means nothing without the number of days.
- The process. How many rounds, roughly how long, what each one is.
- The team and the work. Two or three sentences a candidate would find genuinely informative.
- The awkward ones. Why the role is open. Whether it is a replacement. What happened to the last person. Deciding these in advance is much better than discovering the agent's improvisation afterwards.
The instruction that matters most
Alongside the content, the prompt needs one rule: answer only from the knowledge base; if it is not there, say a recruiter will follow up.
Without that line, the model treats the knowledge base as helpful context rather than a boundary, and will happily supplement it from general knowledge — which is precisely the behaviour you are trying to prevent.
What it does not fix
It is not a negotiation tool. The agent should state a range, not defend it. "Can you do 5 lakhs more?" is a human conversation, and the right behaviour is to note the question and hand it to a recruiter.
It does not make a bad prompt good. If the agent talks too much or repeats itself, more knowledge makes it talk more. Fix the behaviour rules first.
It is not a substitute for saying "I don't know". The most trustworthy agent is one that answers eight questions confidently and defers the ninth. Candidates read that as competence; they read confident wrongness as a bad employer.
The test
Run a test call and be a difficult candidate. Ask about salary, then about remote work, then something you deliberately left out of the knowledge base entirely.
The first two should get real answers. The third should get a clean, unembarrassed "I don't have that detail, but the recruiter will cover it". If it gets an invented answer instead, tighten the instruction before you run the batch — not after.