Plant knowledge assistant
Plain-English questions, cited answers from SOPs, maintenance manuals and quality reports in seconds.
- Where
- Mercedes-Benz
- When
- 2024 – now
- Context
- Prototyped first at Veeco
- Type
- GenAI
200+engineers and technicians using it
85%answer acceptance
−40%time-to-answer in the Veeco pilot
How it works
- 1SourcesSOPs, manuals, quality reports
- 2EmbedChunked and embedded
- 3RetrieveVector search, top passages
- 4GenerateLLM answers from passages
- 5CiteAnswer with sources
The problem
The answers technicians needed were spread across SOPs, maintenance manuals and quality reports. Finding the right paragraph meant searching shared drives or asking whoever knew.
What I built
- Prototyped a RAG chatbot with LlamaIndex over Veeco service manuals and troubleshooting logs.
- Built the plant version at Mercedes-Benz with LangChain and vector search over SOPs, manuals and quality reports.
- Every answer cites its source documents, so people can check it before acting.
- Tracked answer acceptance with users as the main quality signal.
Stack
LangChainLlamaIndexVector searchEmbeddingsAzure OpenAILLM evaluationPython