New Foundations: Organize configuration
Article 2/4
This is the second in the New Foundations series of articles (read the first here) where we share our explorations into the state of AI and how it can be applied to manufacturing and complex system assembly.
At PCH, we’ve built many kinds of software to support product development processes — VR-based automotive package management tools, research and design tools, IIoT asset management platforms, and more.
Over the past two years, we’ve been working to augment the configure-price-quote (CPQ) process for a global producer of large-scale industrial machines. We built a scaleable enterprise AI software solution that allows employees to quickly extract key values from long-document customer specifications. These extracted values are then used to define CAD configuration parameters and a downstream quotation process. The result: improved process efficiency and happier employees.
In this article, we share some of our learnings from the development of LLM-enabled enterprise solutions and our explorations of the evolving space. Explorations driven by the question: How can AI systems give us more clarity on what to build?
Machine configuration requires context-rich communication.
Complex manufacturing is integral to human progress. It’s a process of applying our knowledge to transform raw materials into machines that serve us. Machines that let us travel across the world, see into the body, and harness the power of the sun…
Core to the process of designing, building, and using these machines is context-rich communication. Filtering, reading, comprehending, transforming and sharing facts and ideas related to complicated requirements, systems, and implications of these systems.
Furthermore, many complex machines are, by default, required to be configurable to a certain extent. They’re so big and expensive that to properly install and apply their utility you need to adjust their design slightly to enable a smooth integration.
The point: configuration of complex machines requires lots of information processing. Regardless of the specific machine or configuration workflows, the general ‘sub-tasks’ that eat time and add strain to employees in document processing could be defined as:
- Answering specific questions related to large volumes of reference documents
- Guessing the most likely answer, given the learned context
- Summarizing long texts and complex descriptions
- Translating between languages (technical, cultural)
These are excellent sub-tasks for machines.
Building semantic search solutions
Keyword search is a common method for search. But it is limited for large context-rich data sets. Incorporating large language models (LLMs) into your search can significantly enhance the user experience by allowing users to ask questions and find information in a much easier way.
We built the initial version of this solution around fine-tuned versions of early LLM models (e.g. BERT, Big Bird). The dramatic evolution of the space over the past 18 months has enabled an entirely new tech stack and accessible tools for building these types of systems. From an ecosystem of model APIs (Open AI, HuggingFace, Cohere) and orchestration frameworks (Langchain, LlamaIndex etc.) to vector databases (ChromaDB, Pinecone etc.)
These tools drastically simplify the process of building a semantic search system. They enable the improvement of search accuracy by understanding the searcher’s intent and the contextual meaning of terms as they appear in the searchable data space to generate more relevant results— whether on the web or within a closed system.
To explore this technique for semantically answering questions from vector databases, we built a series of Streamlit mini-apps that all leverage this general architecture and provide examples of what LLMs can do!
- 🔎 Research Assistant 🪄: Ask a research question, get a thread of actionable article summaries source from filtered Google searches.
- 🎙️ Keynote Talk Generator: Ask a question, get a tight keynote presentation outline with topics summaries from Wikipedia.
- 💬 Chat with Reports: Upload PDF reports and chat with them to better understand their context.
- 🖼️ Utopian Narratives 🗣️: Upload a picture, get a short utopian audio story inspired by the scene.
Check out the code repo of some of our most recent experiments into this new tech stack. (Feel free to copy and play with this playground yourself!)
A new era of human-to-machine experience.
We’ve learned that, as the technology to build bespoke LLM solutions becomes commoditized, the greatest challenge becomes the development of an intuitive user experience.
In the past, humans had to speak the language of machines. We had to conform human ideas into rigid data structures. We were forced into becoming more machine-like in our digital interaction.
In the very near future, machines will understand our language. This will revolutionize what it means to have a digital experience. Interaction will be about conversation with machines.
Conversational AI will create incredible opportunity to refine what work means across industries and help us create industrial efficiency in balance with human well-being.
Words by Timo Gmeiner. PCH Innovations is a Berlin-based, creative engineering studio for exploratory technology and innovation strategy.
We are always open to finding curious, talented, and motivated people to help us translate cutting-edge AI research into meaningful applications and products. Does this sound like you? Get in touch!
