Enterprise HR RAG Platform

A live demo is available on request.

What it is

Most “chat with your documents” demos fall apart in an enterprise HR setting for two reasons. First, they hallucinate — and confidently wrong HR policy is a liability, not a feature. Second, they only handle generic questions (“what’s the leave policy”) and have no safe way to answer personal ones (“how much leave do I have”), because that requires reaching into employee data without leaking it across users.

This platform was built to solve both. It’s a production-grade retrieval system that answers HR policy questions and personalized employee questions, always with sources, with the security and privacy controls a real deployment requires.

The core design decision

The heart of the system is hybrid retrieval with an intent router.

Pure vector (semantic) search is good at meaning but misses exact policy terminology. Pure keyword search (BM25) catches exact terms but misses meaning. So the system runs both and fuses the results with Reciprocal Rank Fusion — a method that combines two rankings without needing their scores on the same scale. You get exact-term precision and semantic recall together.

In front of that sits a query router that detects intent. A policy question goes to document retrieval. A personalized question (“my leave balance”) goes to the Cloud SQL HR database. A hybrid question draws from both. This router is what separates a generic document bot from a system that can safely answer employee-specific questions — and it’s where the per-user data isolation is enforced.

Why the evaluation matters

It’s easy to build a RAG demo that looks good. It’s hard to know whether it’s actually accurate — and in HR, accuracy is the whole point. So answer quality is measured with RAGAS across a fixed set of 60 questions. The build improved relevancy from 0.674 to 0.857 with 0.967 source accuracy, and — the part I care about most — that score is wired into CI/CD as a quality gate. Any change that drops answer quality below 0.80 fails the build and never ships. The system can’t silently regress.

Analytics without the privacy cost

Understanding how the system is used — which categories of questions come up, where answers fall short — is operationally valuable, but in HR it’s also a privacy minefield. The solution is to log to BigQuery while storing only SHA-256-hashed identifiers and category metadata: never the PII, never the question text, never the answer. You get the usage analytics that let you improve the system, with nothing that could expose an individual employee. Privacy-by-design and operational visibility aren’t in tension when you log the right thing.

Enhancements along the way

A few extensions beyond the base retrieval system, each driven by what the RAGAS scores and real usage exposed: handling images and tables in source documents (not just plain text), tuning the balance between keyword and semantic retrieval, and tuning chunk size — the unglamorous, measurable work that moves answer quality. The deeper methodology behind these is its own topic; here it’s enough to say the tuning was driven by evaluation, not guesswork.

Beyond HR

Nothing about the hybrid-retrieval-plus-router pattern is specific to HR. The same architecture — fuse keyword and semantic search, route by intent, enforce quality in CI/CD, keep analytics PII-free — is a reference design for enterprise search across any department. HR was the first, well-scoped problem to prove it on.