In LangGraph, knowledge retrieval usually follows a retrieval-augmented generation (RAG) pattern. When a node needs external facts, it creates an embedding of its query, sends it to a vector database, and obtains the top-k similar results. Those results—text snippets, JSON blocks, or prior outputs—are then concatenated or summarized and fed into the next reasoning node.
This separation of retrieval and reasoning lets developers tune each stage independently. You can adjust the number of results, similarity thresholds, or ranking models without touching the rest of the workflow. LangGraph tracks data lineage automatically, so every downstream node knows which retrieval produced its input.
Using Milvus for the retrieval layer keeps this process efficient. Its approximate-nearest-neighbor indexes (HNSW, IVF, DiskANN) handle billions of embeddings with sub-second latency. Developers can shard large collections by project or agent type, stream new embeddings during graph execution, and still maintain consistent recall. The result is responsive, traceable knowledge access inside complex agent graphs.