Building SmithDB, a Data Layer for Agent Observability

LangChain has launched SmithDB, a distributed database specialized for Agent Observability. This database currently powers LangSmith's core workloads.

As AI agents have become more advanced, the volume and size of trace data has exploded. Modern agent traces contain hundreds of nested Spans, and the data arrives in fragmented form, making it difficult for existing observability stores to handle.

SmithDB supports the following features to address these challenges:

• Random access: Instantly load individual runs or traces
• Full-text search: Search phrases and patterns within inputs and outputs
• Tree-aware queries: Filter based on a trace's root and child nodes
• Thread reconstruction: Reconstruct threads for long-running conversations
• JSON filtering: Query custom metadata and structured tool outputs

Built in Rust, SmithDB leverages the Apache DataFusion query engine and the Vortex file toolkit, and consists of object storage, a Postgres metastore, and stateless services. This has improved LangSmith's core experience by up to 12x faster than before.