Agentic Search

Synopsis

Agentic search is retrieval run as an active investigation rather than a single lookup. In the World's Fair material, it shows up as the move away from "vector search once and stuff it into context" toward systems that plan queries, combine retrieval modes, inspect source quality, follow missing evidence, and keep a decision trace of what was searched and why.

Origin And Context

The connected sessions frame agentic search as a convergence of RAG, hybrid retrieval, semantic code search, SQL, UI telemetry, long-context caching, knowledge graphs, web data, and agent evaluation. Turbopuffer's RAG talk argues that retrieval is broader than vector search; Ogilvy's hybrid RAG session adds SQL RRF and UI telemetry; Tesco and Turbopuffer connect the pattern to local code indexes and semantic code retrieval; Neo4j emphasizes decision traces over document dumps; Bright Data focuses on whether an agent actually searched the web in a verifiable way.

Why It Matters

The linked talks repeatedly point to the same failure mode: agents sound confident while retrieving the wrong evidence, ignoring user signals, overloading the context window, or claiming to have searched without a durable search trail. Agentic search matters here because it turns retrieval into an observable workflow: query choices, source triage, reranking, citations, gaps, conflicts, and stopping criteria can be evaluated instead of hidden inside one prompt.

How To Use It

Treat search as a loop with state. Start from the user question and available signals, retrieve across the right surfaces, rerank by task relevance, inspect primary or authoritative sources, record claims with citations, and branch when evidence conflicts. Use BM25 or exact search for identifiers and rare terms, vector search for semantic recall, SQL for structured facts, knowledge graphs for relationships, UI telemetry for product context, and cache or long-context strategies only when they improve evidence quality rather than merely adding bulk.

Where It Is Useful

Agentic search is useful anywhere an answer depends on evidence scattered across multiple systems: research agents comparing sources, enterprise assistants over internal knowledge bases, coding agents searching repositories, compliance workflows correlating documents, product agents turning user signals into pull requests, and support systems that need to explain which logs, tickets, docs, or web pages informed the response.

When To Use It

Use agentic search when the question is underspecified, evidence is distributed, freshness matters, exact citations are required, or the agent must reconcile conflicting sources. A direct database query, deterministic API call, or simple keyword lookup is still better when the answer lives in one known structured source and the task does not require exploration.

Active Use Cases

Related Slide Decks

Related Scheduled Sessions

Related People

Related Companies

Transcript And Resource Support

Transcript-backed resources

Quote signals