Diese Case Study ist auch auf Deutsch verfügbar
TL;DR
The Challenge
Sprengnetter Books' traditional keyword-based search could find relevant documents, but it couldn’t generate direct answers to specific questions. Users had to manually sift through results, which was especially inefficient for technical or fuzzy, vibe-based queries. The goal was to develop a solution that could understand real estate valuation knowledge and provide users with precise, natural language answers.
Our Solution
INNOQ implemented a Retrieval-Augmented Generation (RAG) assistant that combines the modern generative power of the GPT-4o Large Language Model with targeted information retrieval. The solution was developed by a small INNOQ team in just a few months.
The Result
An AI-powered chat system now provides subscribers with natural language access to Sprengnetter Books, going far beyond the capabilities of traditional search functions.
Challenge: Providing Answers, Not Just Search Results
Real estate valuation demands in-depth expertise across many areas—ranging from building regulations to valuation methods and market analyses. Sprengnetter makes this knowledge accessible through its digital database, “Sprengnetter Books.” While the existing full-text search could guide users to relevant documents, it lacked the ability to understand and independently formulate answers. Users had to go through the documents themselves—an often time-consuming task, especially when faced with specific technical questions.
Approach: Agile and Focused Implementation
Working closely with Sprengnetter, INNOQ first analyzed the specific requirements and developed a prototype to demonstrate the capabilities of modern generative AI. A small, dedicated development team then worked in short iterations to implement the solution. This agile approach also provided Sprengnetter with a straightforward and hands-on introduction to generative AI.
The collaboration with INNOQ was highly focused from the start. The team not only brought technical expertise but also developed a deep understanding of how our users work with technical information.
Tina UhligProduct Owner Sprengnetter Books
Technical Approach: State-of-the-Art AI Meets Professional Knowledge Base
The implemented system is built on the RAG architecture (Retrieval-Augmented Generation), which combines the power of large language models with precise information retrieval. Using OpenAI GPT-4o via API and MongoDB Atlas for vector search and PDF storage, the solution delivers real value to users: precise, source-backed answers to technical questions. As one of the leading “world models,” GPT-4o has the necessary generalization capabilities to address a wide range of questions in combination with the retrieved sources. Furthermore, hallucination rates are significantly lower compared to smaller models.
RAG architecture (Retrieval-Augmented Generation) addresses a core issue with using large language models: the tendency to “invent” facts when they encounter knowledge gaps. Instead of making up information, the system first retrieves relevant documents to extract matching information. Only these verified details serve as context for generating answers. The result is clear: the system can harness the language model’s expressive power while remaining grounded in verifiable knowledge.
This architecture has distinct advantages. Unlike a simple language model or traditional full-text search, users receive precise, source-based answers in natural language. RAG thus blends the precision of document-based search with the AI’s ability to understand relationships and convey them clearly.
For a quick technical introduction to RAG architecture, check out our free primer.
One challenge with large language models is their tendency to “hallucinate,” when working around gaps in the knowledge providing factually incorrect but plausible-sounding statements. Our solution mitigates this issue by strictly requiring that every answer generated must be supported by specific references from Sprengnetter Books. This ensures that the system delivers verifiable, technically sound answers.
Special attention was also given to preparing technical information in a context-sensitive way: the system not only understands user questions but can specifically extract relevant details from the knowledge base and present them comprehensibly—even for complex mathematical formulas.
With the new AI-powered assistant, we’ve achieved a quantum leap in the usability of our digital knowledge database. Our customers can now engage in genuine dialogue with our expertise.
Andreas KadlerCEO Sprengnetter Real Estate
Technologies Used
- RAG architecture (Retrieval-Augmented Generation)
- OpenAI GPT-4o via Microsoft Azure
- MongoDB Atlas for vector and full-text search
- PDF Processing Pipeline using PDFMiner, PDF.js Express Viewer
