Purpose-Built Agentic Search Models
Definition
Purpose-built agentic search models are LLMs trained specifically on the search task — the rewrite / retrieve / rerank loop — rather than general-purpose models (GPT-5, Sonnet, Gemini) that happen to be able to call a search tool. Because they are specialized, they can be smaller, faster, and cheaper while matching or beating frontier models on domain retrieval.
The Thesis
Frontier models handle the “80% case” of search well but miss the domain-specific last 20% (e.g. “bistro tables” means “small outdoor tables” in a furniture store, not restaurant equipment). They also conceive of “search” as near-flawless web search, whereas most teams run smaller, focused backends. A model trained on document search in a given domain orchestrates simple retrieval primitives (BM25, an embedding model, a few filters) and supplies that missing domain knowledge.
See Doug Turnbull’s framing in Agentic search models: agentic search unbundles the thick retrieval monolith — query classification, multiple backends, reranking — into thin tool wrappers orchestrated by one model that sees the whole problem.
Mechanism
These models run a multi-turn tool loop:
- Write several query variants (query rewriting for recall)
- Execute them against a search backend
- Pick the best results, ignore noise
- Iterate, then run a dedicated rerank turn
A specialized model converges in 2–3 turns where a general agent may take 7–8.
Examples
| Model | Maker | Notes |
|---|---|---|
| SID-1 | SID.ai | First mover; OpenAI-compatible API; rewrite/retrieve/rerank |
| Waldo | Glean | Enterprise search model |
| (corpus-tailored) | Charcoal | Tailors a model to your corpus |
Trade-offs
Advantages: smaller/faster/cheaper than frontier models; domain-aware; drop-in over an existing backend; reduces the context-engineering and bespoke-pipeline burden.
Limitations: today often too slow for high-QPS site search; quality depends on training data matching the deployment domain; an emerging, fast-moving category.
Related Concepts
- Agentic Search — the multi-turn paradigm these models embody
- Reranking — the terminal step these models specialize in
- Direct Corpus Interaction — an alternative frontier direction: let the agent search raw text directly instead of training a retrieval model
- Query Understanding — query rewriting is a core competency
- RAG — these models can replace or sit alongside RAG pipelines
Articles
- Agentic search models — Doug Turnbull; the conceptual case
- Agentic Search Models with OpenSearch and Elasticsearch — Bonsai; hands-on SID-1 integration