AboutServicesPricingPhilosophyProductsWorkBlogCareersParloirJapanese
← AI CASES

ai-cases

Enterprise RAG for a Food Manufacturer — Scattered SharePoint, Answered with Sources

Built an enterprise RAG system for a food manufacturer. Company rules, notices, financial reports, and internal postings scattered across SharePoint sites are now answered instantly, with sources.

AIRAGazuresharepointknowledge-searchllm

Overview

Everything lived in SharePoint. And still, nobody could find anything.

Company rules sat in document libraries. Notices lived on site pages. Internal job postings were buried in a bulletin board built as a custom list. Whether you found the document you needed depended on knowing someone who knew where it was. The information existed — it just didn’t function as organizational knowledge.

We didn’t build a search box. We built a state where “ask here, and the company answers you, with sources.”

Approach

  1. Unifying structurally different sources into one search foundation. We designed ingestion pipelines per source type — document libraries, site pages, and list-based bulletin boards — across multiple SharePoint sites, converging into a single RAG endpoint on Azure AI Search + Azure OpenAI.

  2. Automation where ingestion gaps are structurally impossible. New folders and new posts are detected by periodic reconciliation jobs, and nothing becomes searchable until registration completes. Zero manual registration work, and zero risk of unregistered content leaking into results.

  3. Source-grounded answers with sensitivity controls. Every answer carries its citations. Sensitive sources such as HR-related materials are display-controlled — the system can answer while withholding where the answer came from, as an enforced rule rather than a convention.

  4. Search quality forged in operation. Ask for “the latest financial report” and an older year comes back — we traced this class of failure from user feedback down to measured root cause, then fixed it structurally by detecting temporal phrasing and blending date-ordered retrieval. LLM model upgrades roll forward and back with a single configuration switch, so improvement keeps cycling inside monthly maintenance.

Outcome

  • Company rules, notices, financial reports, and internal postings — every employee gets source-grounded answers from a single entry point
  • New documents and posts become searchable automatically, with zero manual operations work
  • A feedback loop that traces reports to measured root causes runs inside monthly maintenance. “I couldn’t find it” turns into a structural fix the same day
  • Continuing into the next phase: expanded coverage and search analytics