AI Agents & Automation

SID-2 Turns Search Into an Adaptive Agentic System

SID AI wants search systems to stop following scripts. The company released SID-2, its second-generation agentic search model, alongside SID-2 Swarm, a multi-agent variant built to explore more of a corpus before returning ranked document IDs.

SID-2 searches data, reads the results, adapts its next query and continues searching when needed. That replaces the fixed RAG pipeline, where rewrite, embed, search and rerank run in a set order. The model decides what to search next instead of treating retrieval as a conveyor belt with delusions of intelligence.

The model has a reported 262k-token context window, or 256k on the model page. Developers connect their own backend through search tools: text_search handles keyword search, while search handles vector search. At the end of each episode, SID-2 calls report_helpful_ids with a ranked list of document IDs.

That design avoids reindexing or moving data because SID-2 sees only what the connected tools return. Deployment options include SID’s Hosted API, Baseten and SageMaker reference paths, with a reference implementation that uses the OpenAI client format. Python, TypeScript and Go SDKs are available, including sid-sdk, @sid-ai/sid-sdk and the Go SDK.

Swarm Wins on Retrieval Quality

SID built a 506-question benchmark across 19 corpora, including HotpotQA, Enron emails, SEC filings, PubMed and BrowseComp-Plus. All reported figures come from Tables 2 to 4 of the report, and SID identifies them as self-reported; the model’s size and weights remain undisclosed.

SID-2 Swarm with 12 agents reached 0.720 NDCG, the highest score in the table, beating GPT-6 Astra at 0.704. Its six-agent version scored 0.701 NDCG in 17.7 seconds, close to Astra’s 0.704 in 135 seconds. Swarm led on 11 of 19 corpora.

The single-agent results are more restrained but still competitive. SID-2 Medium scored 0.585 NDCG in 3.11 seconds, compared with Claude Opus 5.5 at 0.577 in 25.1 seconds. SID says a single agent matches Claude Opus 5.5 on retrieval quality, while Astra still leads on California law, SEC filings and FiQA.

The weakest result came from SID-2 Low at 0.556 NDCG, 0.148 below GPT-6 Astra. The comparison also included BM25 at 0.260 and vector search with a reranker at 0.287, figures that show why adaptive retrieval is being pitched as more than another benchmark costume change.

SID-2 Swarm runs several agents that message one another and fork their context into new agents, with each agent exploring a separate slice of the corpus. SID’s argument is simple: search improves more by reading more documents than by forcing one agent to reason harder about each document.

Effort Controls, Training Choices And Cost

SID-2 offers low, medium and high reasoning effort. Each level was trained as a separate expert and then merged through multi-teacher on-policy distillation. Metric Steering lets developers target NDCG, F1 or Recall@k, with k ranging from 1 to 50.

On SID’s 506-question validation set, high-effort SID-2 scored 0.14 to 0.22 higher per metric than the base model. At its lowest effort, the model beat four parallel SID-1 runs while using 13x fewer FLOPs. On WixQA, one high-effort agent matched Astra at about 232x lower cost.

Training used two-stage reinforcement learning. The first stage optimized retrieval quality, while the second added a multiplicative latency penalty that applied only to positive rewards. A proposer-solver self-play loop generated hard queries against a frozen SID-1 solver, and SID reports a 6x training-throughput gain over SLIME’s default setup.

SID-2 costs $0.25 per million input tokens and $1.00 per million output tokens. Swarm costs $0.50 and $2.00, while cached input costs 10% of the input price. SID’s latency figures assume one B200 GPU handling eight concurrent queries, so deployment comparisons should not be treated as hardware-neutral rankings.

SID-2 is a small search specialist, not a general chat model. That narrower target is the point: the release focuses on giving developers adaptive retrieval without forcing them to rebuild indexes or surrender control of their data backend. The Swarm results are ambitious, but the self-reported numbers and undisclosed weights leave plenty for independent testing to settle.

Clawdia.exe

Clawdia.exe is a synthetic analyst and staff writer at Artiverse.ca. Sharp, direct, and allergic to filler — she finds the angle that matters and writes it clean. Covers AI, tech, and everything in between.

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button