Dienstleistungsfinder KI-Suche lets citizens find municipal services on muenchen.de with everyday language. Based on a Retrieval-Augmented-Generation (RAG) approach it understands questions like "Meine Oma braucht einen Rechner, gibt es einen Zuschuss?" and answers with grounded quotes from the official service pages.
The search combines classic BM25 keyword search with semantic embeddings in Qdrant (hybrid search). Results are re-ranked by a separate reranking model and boosted by current page popularity, so that frequently visited services appear higher. For each retrieved document a large language model generates an answer, a literal quotation and a suitability judgment — documents that cannot answer the question are filtered out. The system is evaluated continuously against a human-created gold standard (MRR ≈ 0.67, Recall@10 ≈ 86 %).
The backend additionally exposes the retrieval endpoint as a MCP server, so other AI applications of the city — for example MUCGPT — can query official Munich service information as a context source.
Technical details
- Frontend: Vue web component that can be embedded on any city website.
- Backend: FastAPI service in Python with LangChain for the RAG pipeline and Qdrant hybrid search.
- Indexer: batch job reading service pages from the CMS, transforming them to Markdown and upserting embeddings into Qdrant.
- Models: OpenAI-compatible endpoints for embeddings (text-embedding-3-large) and answers (gpt-4.1-mini) plus a separate reranking model.
- Observability: tracing and prompt management with Langfuse.
More information: https://ki.muenchen.de/ki-systeme/dlf