Agentic RAG Help Hub for Enterprise Retail Operations
Challenges
Finding answers across large, complex SOPs was slow.
Traditional search struggled with business terms and vague queries.
The solution needed to scale to 60,000 users with reliable answers.
Outcome
2X productivity for store colleagues by reducing time spent searching for answers.
95–96% accuracy on the business-provided golden dataset.
60,000 users supported through an enterprise-scale Help Hub.
Solution
Enterprise Knowledge Assistant (RAG)
Challenges
Solution
Technology Stack
Outcomes
For a large enterprise retailer, the answer to many day to day store problems already existed. But the challenge was finding those answers. Store colleagues relied on extensive Standard Operating Procedure (SOP) documents covering around 15 to 16 business modules. These included areas such as food, bakery, clothing and cafés. When something went wrong, from a freezer failure to a till issue, employees often had to search lengthy SOPs, ask a manager or raise a service desk ticket.
Cloudaeon worked with the organisation to change that experience. We built an agentic RAG-based Help Hub on Databricks that allowed store colleagues to ask questions in natural language and receive a relevant, grounded answer within seconds. The solution understood business terminology, handled ambiguous questions, searched across multiple business areas and gave users the original source reference for verification. The platform was designed for an enterprise user base of around 60,000 people.
Challenges
Finding one answer inside large SOP libraries
The retailer had operational procedures across 15 to 16 business modules. Each module contained different processes, terminology and operating instructions. The information existed, but getting to the exact answer was difficult. Store colleagues often had to know which section or category to search before they could find what they needed. For a new employee, this was even harder.
Minor store issues could take hours to resolve
A simple operational problem could follow a long support path. A store colleague would first search the SOP documentation. If they could not find the answer, they would ask their manager. If the issue was still unresolved, it could move to the service desk. This process could take around three to four hours for an issue whose resolution might already be documented.
Business language made traditional search unreliable
Store operations used internal terminology, abbreviations and business-specific phrases. An employee might know the phrase used on the shop floor without knowing how the same information was written inside the SOP. A simple keyword search would therefore not be enough. The system had to understand context and find the closest relevant information rather than match words literally.
SOPs contained more than plain text
The documentation contained text, tables, images, workflows and visual instructions. Some images carried important operational meaning, including procedures, warnings and do's and don'ts. Extracting only the text layer would have removed part of the business context. The system therefore needed to understand different content types without breaking their relationships.
Queries could be vague across business modules
A question could have different meanings depending on the business area. Users did not always specify the module when asking a question. Returning the first available answer created a risk of giving a technically valid but contextually incorrect response. The solution needed to know when it had enough information to answer and when it needed clarification.
The solution had to work at enterprise scale
This was not being designed for a small group of test users. The retailer had an overall user base of around 60,000 people. The architecture therefore had to maintain predictable response performance as adoption increased. The business also needed visibility into how the system was being used, where it was failing and what needed improvement.
Root Cause Analysis
Cloudaeon did not start by choosing an LLM or building a chatbot. We first looked at how store colleagues were actually trying to solve operational problems. We analysed the full information path and discovered that the knowledge was fragmented across large documents. Important context was embedded inside images and tables. Traditional keyword search could not reliably understand store terminology. Users did not always provide enough context in their first question. And a generated answer would not be trusted unless it could be traced back to the original SOP. Instead of building a generic search interface, we designed an agentic RAG system that could retrieve, reason, clarify, verify and return source-backed answers.
Solution
Cloudaeon built an enterprise Help Hub using an agentic RAG architecture on Databricks.
Store colleagues could enter a question through a simple user interface. The system searched the retailer's SOP knowledge base, understood the business context and returned the most relevant response.
How We Delivered
Converted existing operational content into a usable knowledge source: The original operational content was available through enterprise web pages. We used scraping methods to convert the content into PDF files that could be processed consistently through the ingestion pipeline. The challenge was that these PDFs were not simple text documents. They contained images, tables and operational workflows that carried important context.
Built a multimodal PDF parsing process: Rather than relying on a high-cost document intelligence platform for the complete parsing workflow, we used native Python-based processing. This helped retain the original document context during later retrieval.
Created a metadata-aware vector knowledge base: Once the content had been extracted, it was divided into manageable chunks. Vector embeddings were generated for those chunks and stored in the Databricks vector database. Metadata was stored alongside the content. This included information such as the source document and business module. That metadata later became important both for retrieval and for showing users the original source behind an answer.
Added agentic retrieval instead of basic vector search: The retrieval layer used hybrid search. But retrieval alone was not enough. The agent was designed to decide whether the information it retrieved was sufficient to answer the question confidently.
Built a structured evaluation process: The RAG application was evaluated before rollout using both business-provided and synthetically generated test data. The business supplied a golden dataset against which the system achieved approximately 95% to 96% accuracy. We also generated synthetic question and answer patterns from the wider document set. These were evaluated through MLflow using an LLM-as-a-Judge approach.
Technology Stack
Databricks
MLflow evaluation
Agentic RAG
Vector embeddings
Python
OCR
LLM-as-a-Judge
Outcome
The biggest change was the speed at which store colleagues could get operational answers.
2X productivity for store colleagues, particularly by reducing time lost searching for information during normal store operations.
95–96% accuracy on the business-provided golden dataset.
Help Hub also improved trust in AI-generated responses. Every answer remained tied to the retailer's own SOP knowledge base, with the source available for verification.
For the business, Help Hub turned static operational documentation into knowledge that store teams could use while the problem was actually happening.
Conclusion
The retailer did not need more documentation. It needed a better way to use the knowledge it already had. Cloudaeon transformed a large, complex SOP estate into an agentic Help Hub that could understand context, ask for clarification, retrieve across multiple business areas and return answers that employees could verify. If your teams are still spending hours searching SOPs, manuals and internal knowledge before they can act, Cloudaeon can help turn that information into a governed, enterprise-ready knowledge assistant. Talk to an expert now.