Tens of thousands of PDFs scattered across the company intranet, and you want AI to accurately answer "what's our company's reimbursement policy" — this week's Juejin piece, Hands-on with an Enterprise Knowledge Base Q&A Agent, tells us: the real time sink isn't the LLM. It's the dirty work of document parsing, vectorization (turning text into numbers AI can "understand"), and retrieval optimization.
What this is
The scenario: a company's intranet holds tens of thousands of PDFs (policies, procedures, project docs), and employees want AI-powered Q&A. The technical approach is today's mainstream RAG (have AI first search relevant passages from the document store, then answer based on them — avoiding fabrication from thin air).
The genuinely worth-dissecting piece is its "anti-hallucination trio": RAG retrieval supplies the factual foundation; "faithfulness self-check" — using AI to check AI, where every assertion must trace back to the source text; and "reflective re-answering" — if AI finds its answer uncertain, it re-searches; if nothing turns up, it explicitly says "not found," no faking. On top of that, MCP (a protocol standard letting AI plug into enterprise internal systems) bridges OA and HR — so AI doesn't just tell you "what the policy is," it can also tell you "how to file a reimbursement."
Industry view
Supporters see this as a template for enterprise AI deployment: over the past two years, everyone's been racing on LLM parameters and running benchmarks, but what actually decides whether the thing is usable — whether anyone dares to use it — is the dirty work: document processing quality, retrieval accuracy, answer traceability. The article's "metadata trio" (source, dept, type) lets retrieval be precise down to the department level — an engineering detail that's rarely explained this clearly.
But the dissenting voices are sharp. One enterprise IT lead commented bluntly: real-world deployment has to solve three actual problems first — OCR on scanned documents (recognizing text in images) often underperforms marketing claims on tables and stamps, so the source data is dirty from the start; documents get updated weekly, so who pays for index rebuilds; and most critically — Chinese workplace employees don't accept "not found" as an answer. They want a definitive answer, whether or not it's correct.
Other voices caution: MCP is still in the protocol-layer competition phase. Anthropic pushing it alone doesn't make it an industry standard — betting too early carries lock-in risk.
Impact on regular people
For enterprise IT: Over the next year or two, AI project evaluation criteria will shift from "can we use an LLM" to "can the document pipeline actually run." IT departments need to build data governance capability — not just know cloud and servers.
For individual professionals: The time white-collar workers spend "looking up policies and procedures" will likely be compressed substantially by AI. But flip it — if your job is just "know a rule and tell someone else," you'll be replaced faster than you think.
For the consumer market: No direct short-term impact. But once enterprise-internal AI Q&A stabilizes, the next wave spilling into B2C will be "AI customer service that actually solves your problem" — not the current script-driven bots that miss the point entirely.