AI knowledge assistants for a 30 000-employee grocery chain

The problem was three groups asking the same questions in the wrong place, with a hard rule of no invented answers: an assistant that confidently misquotes a pricing regulation is worse than no assistant. The risk is well documented — in the Stack Overflow Developer Survey (2024) the most cited concern about AI tools was distrust of their output, and OWASP’s LLM Top 10 (2025) lists misinformation and prompt injection among the main application risks.

• Commercial staff spent hours finding the right clause across dozens of regulations; new hires took months to become productive and pulled their managers into answering the same questions.
• Store directors phoned head office with identical operational questions — and could not search a 40-page PDF while standing in the dairy aisle.
• HR answered the same leave, sick-pay and scheduling questions for 30 000 employees, in several languages, over and over.

Every answer is built strictly from retrieved fragments and shows its sources as numbered citations with a deep link into the document. If the corpus does not contain the answer, the assistant says so and points to the responsible expert.

It was built as one retrieval-augmented generation platform with three surfaces. Documents are audited, chunked along clause structure and indexed for hybrid search; the model answers only from retrieved fragments, cites them, and returns an explicit “not found” when retrieval confidence is low. Grounding and logging were treated as governance controls in the sense of the NIST AI Risk Management Framework (2023) — measurable, monitored and owned — not as prompt tricks.

• Document corpus audit first. Before indexing anything we walked every function through its documents and flagged outdated versions — a RAG system on stale documents is confidently wrong.
• Indexing pipeline. Parsing of DOCX/PDF/XLSX, semantic chunking that respects clause structure, multilingual embeddings, hybrid search (BM25 + vector) with a cross-encoder reranker. Re-indexing runs on document change.
• Grounded generation. Answers cite fragments as [1][2]; all prompts and answers are logged.
• Three surfaces, one platform. Web app and Telegram bot with corporate SSO for commercial staff; a mobile voice assistant for store directors (local Whisper speech-to-text, format-aware retrieval, text and voice answer); an HR assistant that answers general policy questions and routes personal cases (“how many leave days do I have left”) to a specialist until the HR-system integration is approved.
• Feedback loop. Thumbs up/down on every answer and a weekly digest of unanswered questions — which became the input for updating the regulations themselves.

The table explains why the platform retrieves documents rather than fine-tuning a model on them. Regulations change weekly and every answer must point to a clause; retrieval handles both, while a fine-tuned model would need retraining on each change and cannot cite. Fine-tuning still has a place — tone, format, domain vocabulary — and can sit on top of retrieval.

Store directors ask by voice from a phone. Speech-to-text runs on-premise; the retrieval step knows the store format, because the procedures for a hypermarket and a convenience store differ.

Measured in the pilot group (commercial function, 180 users; 40 pilot stores) over the first three months, then confirmed on the rollout.

The stack is Python services (LangChain for orchestration, custom retrieval), PostgreSQL with pgvector, Whisper for on-premise speech recognition, Laravel + Vue for the admin and analytics, Telegram Bot API and corporate AD/SSO, deployed in the client’s Kubernetes cluster so that documents and transcripts never leave the corporate perimeter. Model choice is pluggable: the pilot ran on a hosted LLM under a data-processing agreement of the kind GDPR Article 28 requires between controller and processor (2016); the HR assistant, which handles the most personal questions, runs on a self-hosted model.

The same platform now hosts the supplier-correspondence assistant for procurement and a contract-deviation checker for legal. The unanswered-questions digest turned into a standing agenda item for the regulation owners — the assistant became the fastest way to discover which rules are unclear. Beyond retail, the same platform shape serves any organisation whose staff answer questions from a corpus of internal documents: banks, insurers, logistics operators, public bodies.