Agentic search treats retrieval as a stateful reasoning loop with query reformulation, tool calls, and memory across iterations.
2
Composable search tools and predictable backends make agent behavior easier to plan, debug, and evaluate than a large monolithic search API.
3
Production systems can route most queries through cheap reasoning paths while sending ambiguous queries through deeper agentic search.
Summary
Santoshkalyan Rayadhurgam argues that search systems must account for users expressing partial intent instead of complete, well-formed queries. Examples such as finding a Python memory resource from last week require entity detection, temporal reasoning, domain classification, and inference about the missing content type. His proposed answer is agentic search, where retrieval becomes a stateful control loop with query reformulation, strategy selection, multiple retrieval backends, and confidence tracking. He recommends exposing simple, composable tools such as keyword and semantic search rather than hiding everything behind a large endpoint. BM25 can work well with an agent because its scores are predictable. He also describes intent-conditioned query embeddings, hybrid relevance signals, and tiered reasoning costs. In his production example, routing, caching, a stateful orchestrator, hybrid retrieval, and multi-stage ranking operate under latency and zero-result constraints. The discussion covers evaluation through trajectory metrics such as convergence rate, entropy reduction, and information gain per iteration. Rayadhurgam expects search to move toward prediction and eventually toward systems that anticipate information needs.
Real users often provide partial intent rather than complete search instructions
Rayadhurgam says traditional search assumed that users would express a fully formed intent. That assumption fails when people type requests such as "find that Python memory thing from last week" or "a laptop for editing but light." These queries contain an entity, a temporal constraint, a domain, and an unspecified content type. The system must infer whether the user wants an article, tutorial, or code snippet. Keyword matching cannot resolve those missing parts reliably because the problem is interpretation rather than retrieval alone. He describes the resulting failures as zero-session memory, intent collapse, and lexical brittleness.
Agentic search turns retrieval into a stateful reasoning loop
The proposed agentic search state stores the original query, a reformulation trajectory, embeddings for each iteration, retrieval strategies, confidence curves, and diversity metrics. This session-local memory persists across tool calls and iterations. A controller can detect failure signals, update individual components, and choose a new strategy. The loop begins with query understanding, including entity extraction, intent classification, and ambiguity detection. It then selects among lexical, semantic, graph-based, or hybrid retrieval before orchestrating multiple backends with fault tolerance. Rayadhurgam describes this as online optimization rather than a single retrieval endpoint.
Composable tools give agents a more predictable search substrate
Rayadhurgam contrasts a monolithic search API with many parameters against composable tools that expose transparent operations such as keyword search and semantic search. A large endpoint is brittle for language models because its behavior is hard to reason about and its nondeterminism makes debugging difficult. Atomic tools make planning more tractable and give the system clearer policies. He argues that these tools depend on an understandable backend. In his benchmark example, BM25 paired with an agent outperforms a complex neural network used without an agent. Predictable scores help the agent form and test accurate hypotheses about search results.
Query interpretation should preserve multiple hypotheses
Agentic interpretation converts free text into structured meaning that includes query type, semantic intent, temporal constraints, and multiple possible interpretations. Rayadhurgam prefers a probabilistic interpretation space over one forced answer because ambiguity can be handled by downstream retrieval strategies. He also distinguishes intent-conditioned query embeddings from document embeddings. A query about a laptop for coding emphasizes CPU, RAM, and the development environment. A video-editing query emphasizes GPU throughput and display characteristics, while a portable workstation query emphasizes mobility and energy use. He reports roughly a 35% precision lift with limited latency overhead.
Relevance scoring needs behavioral and human signals
An LLM can provide a useful prior for relevance, but Rayadhurgam says it does not personalize well, adapt to drift, or explain its failures. His hybrid relevance approach combines model signals with clicks, dwell time, negative feedback, and human feedback. Clicks and dwell time can be strong but vary with position bias and personalization, so they need normalization. A user saying that a result was not helpful is less frequent but can be highly precise, which may justify giving it disproportionate influence or treating it as a veto. The weighting should therefore adapt to different operating regimes instead of remaining a fixed formula.
Reasoning depth should match query complexity and cost limits
Rayadhurgam describes several reasoning tiers. Cached patterns take about 10 milliseconds and handle deterministic, repeated intents. Distilled models take about 50 milliseconds for simple reasoning. A single-pass agent takes about 200 milliseconds, while full reasoning takes about 500 milliseconds and costs more. The system can keep most traffic on efficient paths and reserve deeper reasoning for complex queries. His production architecture includes query routing, real-time streams, caching, a stateful agent service that could run on Temporal, hybrid lexical and vector retrieval, and multi-stage ranking. He gives a production example with about 100 milliseconds of p5 latency and a 6% zero-result rate.
Evaluating an agentic search system requires more than measuring one retrieval hop. Rayadhurgam recommends examining the full reasoning trajectory. Useful measures include convergence rate, the number of steps before termination, the number and order of strategies explored, and information gain per iteration. Retrieval evaluation measures signal fidelity, while agent evaluation measures policy quality. The combined system can be assessed through joint measures such as convergence rate, entropy reduction, and the quality of the reasoning path. This separation helps identify whether a failure came from poor retrieval signals or from the agent's choice of policy.
Better search may become prediction rather than explicit retrieval
Rayadhurgam expects search to move through several stages. Near-term systems may support multi-turn clarification, cross-session memory, and real-time user learning. Around 2026, he expects more domain-specialized reasoning agents. Longer-term systems may become ambient, multimodal, and available across devices and contexts. He compares this direction with recommendation systems, which shifted from retrieving relevant items to predicting what a user will watch next. Search may similarly predict what a user will ask next. His broader claim is that as systems understand intent, context, and world state, information retrieval could become understanding or even reality querying.
