Why Your RAG Assistant Gives Confident but Wrong Recruitment Timelines

Retrieval-augmented generation works by converting text into vectors, storing those vectors, and returning passages nearest to the user query. That design maps well to documents that change rarely: work instructions, vendor setup sheets, engineering change notices, customer specifications. Those are written once, revised infrequently, and the retrieved passages remain valid between queries.

What happens when a retrieval layer indexes that field? It captures the date at the moment of indexing. Later, when users ask "when will 88213 ship," the system returns that indexed date, wrapped in the same confident prose it uses for stable documents. The output gives no signal that the date is a snapshot and may already be obsolete.

• The system was never built to refuse or qualify answers about live state. It synthesizes and supplies an answer.

• Stale retrieval is a known failure point in RAG deployments and appears high in published case studies of real-world failures.

Where RAG produces tangible operational value

RAG delivers large, fast gains on document-shaped problems. Examples:

• Finding the correct revision of a setup sheet that a planner used to spend hours hunting for.

• Answering whether a part configuration has ever been run, with a citation back to the original routing sheet.

Why "just connect it to the MES" isn't a silver bullet

Practical mitigations you can apply now

• Draw clear data boundaries. Use RAG for document retrieval and a separate, real-time query layer for scheduling fields.

• Implement explicit refusal behavior: the assistant should decline to provide a firm ship date unless it can query the scheduler at the moment of the request.
• Add checks that compare retrieved schedule values against the live scheduler and flag mismatches before returning an answer.
• Train users and stakeholders: explain what the assistant can reliably do and what requires a human or live-system check.

The problem is a category error, not dishonesty. When deployed with clear boundaries and the right mix of retrieval and live queries, RAG yields big productivity wins. Left unbounded, it confidently hands out dates that were true only at indexing time and creates operational risk.