Agentic RAG Platform
A production retrieval layer combining agentic chunking, hybrid vector and graph search, reranking, and automated RAGAS evaluation.
LangChain LangGraph RAGAS Langfuse Azure AI Search Vector Databases Graph Databases Python
Overview
A production retrieval-augmented generation platform serving AI products across Carl Zeiss AG. Rather than a single vector lookup, the system runs a staged component flow — agentic chunking, hybrid retrieval across vector and graph databases, reranking, and automated evaluation — so that grounding quality is measured continuously rather than assumed.
Key Features
- Agentic chunking that prepares source content for retrieval instead of splitting on fixed token windows
- Hybrid search spanning vector databases and graph databases, combined with keyword search and metadata filtering
- Reranking stage that prioritizes retrieved context before it reaches the model
- Continuous RAGAS and Langfuse evaluation covering relevance, groundedness, and response quality
- Reusable grounding path consumed by AI-MCP-connected products
Impact
Achieved 94% retrieval accuracy on an internal benchmark, increased response relevance by 50%, reduced retrieval latency by 30%, and cut production agent errors by 50% through systematic evaluation.