SID models: performance, specifications and pricing for SID-2, SID-2 Swarm and SID-1

Models side by side

SID models perform only one task: search. This specialization allows them to achieve high retrieval quality with substantially less compute than general-purpose models. Read our research.

New SID-2 New New SID-2 Swarm New SID-1
Retrieval quality BetterBestGood
Easy question min. latency 0.5 s9 s1.5 s
Hard question min. latency 1.9 s13 s4 s
Reasoning levels Low, medium, highHighDefault
Context window 256kEffectively infinite32k
Supported objectives NDCG, F1, recall@NNDCG, F1, recall@NNDCG (default)
Sample applications Knowledge bases, customer support, e-commerceLegal, finance, healthcare, compliancePlease use SID-2

How it works

Static search

Most retrieval stacks follow a rigid pipeline: rewrite, embed, search, rerank. Every decision is hard-coded at design time, with no way to account for how hard a question is or what the search returned. Even when the reranker spots bad results, it cannot ask for better ones. As no fixed set of choices will fit every question, static pipelines accumulate a long tail of failures.

Question LLM rewrite optional Search backend Reranker optional Ranked IDs

Agentic search

SID models search iteratively: they pick tools, read the results, reason, and repeat until they find what they need. They return a ranked list of relevant documents, which can be shown to a user, passed through a pipeline, or handed to another model or agent. Read more in our docs.

Question Ranked IDs SID-2 tool calls(BM25, ANN, etc.) content +metadata Search backend

Price

Per 1M tokens
New SID-2 New SID-2 Swarm SID-1
Input $0.25$0.50$0.75
Cached input $0.025$0.05$0.075
Output $1.00$2.00$4.50

SID at work

We build SID for the real world and judge it by what improves for our customers: better answers, faster responses, lower cost. We encourage customers to evaluate SID on their own data. If you’d like help designing an eval, talk to us.

Something else in mind? If you have similar or completely different search problems, we’d love to hear from you.

Talk to us

  • Agent memory After roughly six months of conversations, one personal agent saw memory recall deteriorate for its most active users. Repeatedly compressing memories erases detail available in raw conversations, leading to suboptimal personalization.
  • Recruiting Semantic search frequently recommended jobs that didn’t match candidates’ stated preferences. For example, “software engineer not crypto” previously returned jobs at crypto companies.
  • E-commerce In the case of a marketplace with user-generated content, fixed retrieval frequently missed relevant products because of inconsistent naming and incomplete attributes.
  • Customer support Low-precision fixed retrieval flooded the answer generation agent with irrelevant docs, increasing latency and reducing answer quality for live chat.
  • Healthcare and insurance Fixed retrieval returned semantically relevant but less authoritative sources above trusted medical guidelines, failing to respect a defined credibility hierarchy.
  • Legal Semantic search often returned legally irrelevant documents, such as those from the wrong jurisdiction, an overruled precedent, or a lapsed contract.
  • Finance A financial auditing platform needed to find every transaction associated with an account ID across spreadsheets, emails, and nested attachments. Frontier model sometimes took 30 minutes to complete this task.
  • Procurement and compliance Fixed retrieval struggled with questions like “list all service locations” because no single document contained the full answer. Its reranker discarded relevant evidence scattered across multiple documents.
  • Knowledge bases Enterprise search users expected simple article lookups to return quickly but tolerated higher latency for better results on complex multi-part questions. Fixed retrieval pipelines left them dissatisfied in at least one case.

Try SID-2 today.

SID searches your data where it lives, using your existing search backend. No reindexing needed.