LangSmith added AI Search on October 13, 2026, in an announcement on the LangChain blog. When teams review what their agents have been doing, what they usually want is not a keyword but a kind of behavior: the run where the agent claimed a task was finished when it was not, the user who was clearly getting frustrated, the request in which personal information leaked. Trace data has no ready-made field for such questions, and tags, metadata or full-text search cannot surface them; until now the options were to set up an evaluator in advance or to read traces by hand. AI Search lets users describe the behavior they are looking for in natural language and returns the traces where it actually happened.
Why it does not use vector search
The usual way to search by meaning is embeddings: compress the question and every trace into a short list of numbers and return the closest matches. But a trace carries a huge amount of information, and squeezing it into one vector is lossy — a trace where the agent did the work and one where it only said it did can end up with nearly identical embeddings, exactly the distinction teams care about most. LangChain instead has models read the trace text at search time. AI Search sits on SmithDB, a database built for trace data, which first reshapes records from the OpenAI and Anthropic SDKs, LangChain, LangGraph, the Vercel AI SDK, OpenTelemetry and other sources into a deduplicated session trajectory. Structured filters narrow the candidates, then fast decision models check each remaining trajectory against the description, many in parallel, streaming matches back as they are found.
An 8K budget, filled in priority order
Reading raw text has a cost problem: agent trajectories are long. LangChain sampled 150 conversations from Open SWE, its own coding agent; the median ran to about 72K tokens, of which user messages and agent replies were only about 2K, roughly 85% was tool output such as file reads and shell results, and a few conversations passed a million tokens. So each trajectory gets an 8K token budget, filled in priority order: user messages and the agent's replies to the user first, then intermediate reasoning, tool calls and tool results, with anything that no longer fits trimmed away.
The practical payoff is in evaluation and debugging: building an eval set for a failure type, or finding specific conversations for an annotation queue, no longer requires predicting every question in advance — teams can search for whatever occurs to them afterwards. AI Search is in public beta on LangSmith Cloud in the GCP US, EU and APAC regions at no additional cost, switched on by an organization admin. Its limits are stated plainly: judgments come from a model reading traces and can be wrong, and details cut by the token budget — often in tool results — may hold the answer to other questions. Picking context based on the question, second passes when the model is unsure, and grouping or summarizing results are listed as next steps. Traces only become truly useful for observability when they move from being stored to being question-able.
