AI-powered HRMS with semantic policy search

HR documentation is not missing. It is unfindable. Leave policy, notice periods, expense rules and compliance guidelines all exist, written carefully, and they sit in long PDFs organised by the logic of the person who wrote them rather than the question the employee actually has.

So people do the rational thing and ask a human, and the HR team spends its week answering the same forty questions. Keyword search does not solve this, because employees ask in their own words — someone asking how much notice they have to give is not going to type the phrase the handbook uses.

Semantic retrieval is a genuinely good fit for this shape of problem: a bounded, authoritative, slow-changing corpus where the gap between how the answer is written and how the question is asked is the entire difficulty.

The search layer is built with LangChain over HuggingFace embeddings, indexing policies, handbooks and compliance guidelines so they can be queried conversationally rather than navigated.

Policy text has a useful property that general web content does not: it is structured, deliberately scoped, and written to be authoritative. That makes chunking tractable and makes the boundary of what the system should know unusually clear — a question outside the corpus is a question the system should decline, not guess at.

Answers come back in plain language, with the policy they came from. For HR content this is not a nicety. An employee acting on a paraphrase of a leave policy needs to be able to check the paraphrase, and an HR team needs to be able to see which clause the system read before it said something.

That is the same principle as the grounding checks in the retrieval work: a system that cannot show its source cannot be audited, and a system handling compliance content that cannot be audited should not be shipped.

The search sits inside a working HRMS, not beside it — a React front end over PostgreSQL, cloud-native, with the retrieval layer as one capability of the product rather than a separate tool people have to remember exists.

This is the part that tends to get skipped. Retrieval quality is worth nothing if the answer lives in a demo the employee never opens; the model work only pays off once it is inside the system people already use to book their leave.