Stemar

BACKGROUND

Please describe your company and position.

I am the Founder of a web development studio.

Describe what your company does in a single sentence.

We're a web development studio, specialising on complex solutions. 

OPPORTUNITY / CHALLENGE

What specific goals or objectives did you hire Stemar to accomplish?

• Case adopted RAG approach to the project
• To build an AI-powered knowledge base for a large industrial manufacturing environment

SOLUTION

How did you find Stemar?

Referral

Why did you select Stemar over others?

• Pricing fit our budget
• Great culture fit
• Good value for cost
• Referred to me
• Company values aligned

How many teammates from Stemar were assigned to this project?

2-5 Employees

Describe the scope of work in detail. Please include a summary of key deliverables.

The scope was to build an AI-powered knowledge base for a large industrial manufacturing environment where operators needed fast and reliable access to more than 10,000 pages of technical documentation, including PDF manuals, CAD specifications, and Excel maintenance logs. 

The key deliverables included: 
* A custom parser for technical documentation that preserved the structure of engineering tables, specifications, and maintenance instructions. 
* Table-aware chunking logic, so rows containing part numbers, torque values, equipment parameters, and technical codes were not split incorrectly. 
* A hybrid retrieval system combining dense semantic search and BM25 keyword search. 
* Integration with Azure OpenAI to generate answers based on retrieved source materials. 
* A FastAPI backend for querying the knowledge base. 
* Source references in responses, allowing operators to trace answers back to the original manuals and pages. 
* A Ragas-based evaluation pipeline to monitor retrieval quality and answer reliability. 
* Production-oriented monitoring of key quality metrics such as Faithfulness and Recall. 

The most important part of the project was that Stemar did not use a generic RAG approach. They adapted the retrieval and parsing logic specifically for industrial documentation, where tables, part numbers, technical codes, and diagrams are critical.

RESULTS & FEEDBACK

What were the measurable outcomes from the project that demonstrate progress or success?

The project significantly reduced the time required to find technical information. Previously, operators could spend up to 20 minutes searching through manuals and maintenance documentation. With the new AI knowledge base, they could get relevant answers in around 5 seconds. 

The retrieval quality reached 0.93 based on Faithfulness and Recall evaluation through Ragas. This was important because the system had to provide reliable answers for engineering and maintenance use cases, not just general text search. 

Another important outcome was that operators received answers with precise links to the original source pages, which increased trust in the system and made it easier to validate the information before taking action.

Describe their project management. Did they deliver items on time? How did they respond to your needs?

Stemar managed the project in a structured and transparent way. They broke the work into clear phases: discovery, document analysis, parser development, retrieval implementation, LLM integration, testing, and quality evaluation. 

Communication was clear throughout the project. They explained technical trade-offs in a practical way and helped us understand why a standard RAG setup would not be enough for this type of industrial documentation. Their team was proactive in identifying risks, especially around document structure, table parsing, and retrieval accuracy. 

They were also flexible when we needed to refine the search logic and improve how technical tables and specifications were processed.

Stemar team delivered the project within the agreed fixed-scope timeline of approximately 8–12 weeks. The main milestones were completed on schedule, and the team kept us informed about progress, risks, and dependencies. They communicated early and proposed practical solutions instead of delaying the project.

Stemar responded very well to our needs - they quickly understood that this was not a simple chatbot or document search project. The system had to work with complex industrial documentation, exact part numbers, maintenance instructions, and technical specifications. They adapted the solution to the real use case instead of trying to force a generic AI approach. Overall, their response was practical, technically strong, and focused on production quality.

What was your primary form of communication with Stemar?

• Virtual Meeting
• Email or Messaging App

What did you find most impressive or unique about this company?

The most impressive part was Stemar’s ability to combine AI engineering with a real understanding of industrial documentation. Many AI vendors can build a basic RAG chatbot, but Stemar understood why generic text splitting and semantic search were not enough for engineering manuals, CAD-related specifications, and maintenance logs. 

Their custom parsing approach, hybrid retrieval setup, and quality monitoring pipeline made the solution much more reliable. We also appreciated that they focused on measurable quality, not just a demo. The use of Faithfulness and Recall metrics gave us a clear way to track whether the system was actually improving.

Are there any areas for improvement or something Stemar could have done differently?

Overall, we were satisfied with the collaboration. One possible improvement would be to provide more documentation templates at the beginning of the project, especially for acceptance criteria, test cases, and quality benchmarks. This would make it easier for non-technical stakeholders to follow the evaluation process from day one. 

That said, Stemar was responsive and adjusted quickly once the project moved into testing and validation. The final result met our expectations.