Agent-based Search Models

Existing search was a thick monolith with embeddings, rerankers, query understanding, and BM25 bolted together. Query classification, backend retrieval, post-processing, and reranking each moved separately, and each component couldn't see the full context.

In contrast, agent-based search has an LLM orchestrate the entire search process. The model looks at tools and knowledge together, chooses the retrieval primitives it needs, explores candidates, and even leverages other sub-agents or knowledge bases to produce an answer.

General-purpose models like GPT-5 or the Sonnet family are strong in the 80% zone, such as well-organized web search, but they tend to miss the domain-specificity of the remaining 20%. In furniture search, for example, general models often don't grasp the semantic nuance that "bistro tables" means small outdoor tables rather than restaurant equipment.

Right now, context engineering is layered on top with constraints and validation to fill these gaps, but that costs tokens and money accordingly. This leads to the argument that a dedicated LLM specialized for search is needed.

• SID-1 is presented as a dedicated model that's smaller and has lower latency than GPT-5.
• Waldo is a search-specific model released by Glean.
• Charcoal is a startup that tunes search to fit a corpus.

Currently these start from RAG and chunk-based passage retrieval, but going forward there's a good chance they'll expand into domain-specific agent-based search models for areas like e-commerce or job search. The search stack could become simpler, with only basic keyword search, an embedding model, and a few filters remaining while an agent orchestrates the rest.

It's still too slow to use as the main driver of site search today, but the outlook is that search will be reorganized into a different shape than it is now.