The Business Challenge
The business situation the organisation faced before the project began.
Employees often need to search through multiple HR policy documents to find information about leave, onboarding, benefits, and workplace policies. Manual document search can be time-consuming, while general-purpose LLMs may generate answers that are not grounded in the organisation’s actual policies. The challenge was to provide a faster and more reliable way to access policy information while reducing unsupported or hallucinated responses.
How the Solution Was Delivered
Discovery, design, development, testing and rollout — the journey, not the tooling.
- 1Define HR policy question-answering scope and knowledge boundaries
- 2Collect and clean public HR policy documents
- 3Load and split documents into overlapping text chunks
- 4Generate sentence embeddings for policy content
- 5Build and persist the FAISS vector index
- 6Configure semantic retrieval with top-k relevant chunks
- 7Implement history-aware question reformulation for follow-up queries
- 8Build grounded answer generation with fallback handling
- 9Integrate the RAG workflow into a Streamlit chat interface
- 10Test direct, paraphrased, follow-up, and out-of-scope questions
- 11Document limitations and future production enhancements
Business Outcomes & Impact
The measurable outcomes recorded on this AI Case Study — with the story behind them.
Demonstrated end-to-end retrieval and grounded question answering across direct policy questions, conversational follow-ups, and out-of-scope queries, with explicit fallback behavior when relevant policy information is unavailable.
Business Outcome Categories
The story behind the numbers
Delivered a functional end-to-end conversational RAG prototype for HR policy knowledge access. The solution enables natural-language policy search, retrieves semantically relevant content, supports context-aware follow-up questions, and generates answers grounded in the retrieved documents. It also handles unsupported questions through explicit fallback behavior, demonstrating improved knowledge accessibility, employee self-service, and reduced reliance on manual document search.
Lessons Learned
What surprised the team, what worked well, and what would be done differently.
- RAG quality depends heavily on document quality, chunking strategy, and retrieval relevance, not only on the LLM.
- Grounded prompting helps reduce unsupported answers but cannot compensate for missing or irrelevant retrieved context.
- History-aware question reformulation improves retrieval for ambiguous conversational follow-up questions.
- Data freshness is critical because a vector index only reflects the source documents available at ingestion time.
- Local LLM inference improves privacy and avoids API costs, but performance is constrained by available CPU, GPU, and memory.
- Production RAG systems require systematic evaluation, source citations, observability, security controls, and automated knowledge-base updates.