SID-1
SID-1 is SID.ai’s first purpose-built agentic retrieval model — a small LLM trained specifically on the rewrite / retrieve / rerank loop rather than general-purpose tool use. It is a concrete instance of the Purpose-Built Agentic Search Models category.
Characteristics
- Interface: OpenAI-compatible API (
api.sid-1.com/v1). Tool definitions go in the system prompt as structured text, not the OpenAItoolsfield. - Loop: typically 2–3 retrieval turns plus a dedicated reranking turn (
report_helpful_ids); up to 6–7 turns for long/paragraph queries. - Claimed quality: ~1.9x more likely to surface the right result than embedding-only search; more accurate than agentic retrieval on Gemini 3 Pro, Sonnet 4.5, and GPT-5.1 at highest compute, while ~24x faster.
- Cost: priced similar to gpt-5.4-mini (~$0.00093/query in one Gutenberg-corpus test), cheaper and faster than the larger frontier models it beats.
Why a Small Model Wins
Specialization plus focused training on a single task (search rewrite/retrieve/rerank) lets SID-1 outperform much larger general models that must support every conceivable tool and task.
Articles
- Agentic Search Models with OpenSearch and Elasticsearch — Bonsai’s hands-on integration against OpenSearch
- Agentic search models — Doug Turnbull; positions SID-1 as the first mover