SID AI Releases SID-2 Agentic Search Model With 0.720 NDCG Swarm Mode

SID AI has released SID-2, its second-generation agentic search model, alongside SID-2 Swarm. SID-2 is an agentic search model from SID AI that searches your data, reads results and returns ranked document IDs. It replaces fixed RAG pipelines with a model that decides what to search next.

TL;DR

  • Size: Context window is 262k tokens per the report (256k on sid.ai/models).
  • Runs on: Hosted API at platform.sid.ai, plus Baseten and SageMaker reference paths. SID’s latency figures assume 1 B200 GPU at 8 concurrent queries.
  • Performance: A single agent matches Claude Opus 5.5 on retrieval quality. Swarms beat every model SID tested.
  • Best: SID-2 Swarm (12 agents) scores 0.720 NDCG, above GPT-6 Astra at 0.704.
  • Worst: SID-2 Low scores 0.556 NDCG, 0.148 below GPT-6 Astra. Astra still leads on California law, SEC filings and FiQA.
  • Bottom line: Best: frontier-level retrieval at a fraction of the latency. Worst: no disclosed size, weights or independent evals yet.

What is SID-2 and how does it work?

SID-2 is a small search specialist, not a general chat model. A static pipeline runs rewrite, embed, search and rerank in a fixed order. SID-2 instead searches, reads results, adapts and searches again.

Developers implement the search tools against their own backend. A text_search tool covers keyword search and a search tool covers vector search. The model ends each episode by calling report_helpful_ids with a ranked list. No reindexing or data movement is required, since SID-2 only sees what your tools return.

SID-2 Swarm runs several agents that message each other and fork their context into new agents. Each agent explores a separate slice of the corpus. SID argues search gains more from reading more documents than from reasoning harder about each one.

What are reasoning effort and Metric Steering?

SID-2 offers low, medium and high reasoning effort. Each level was trained as a separate expert, then merged with multi-teacher on-policy distillation. Metric Steering lets users target NDCG, F1 or Recall@k, with k from 1 to 50. On SID’s 506-question validation set, SID-2 at high effort scores 0.14 to 0.22 higher per metric than the base model.

How was SID-2 trained?

Training used 2-stage reinforcement learning. Stage 1 optimizes retrieval quality only. Stage 2 adds a multiplicative latency penalty that applies only to positive rewards. A proposer-solver self-play loop generated hard queries against a frozen SID-1 solver. SID reports a 6x training-throughput gain over SLIME’s default setup.

How does SID-2 perform on benchmarks?

SID built a 506-question benchmark across 19 corpora. These include HotpotQA, Enron emails, SEC filings, PubMed and BrowseComp-Plus. All numbers below are self-reported from Tables 2 to 4 of the report.

  • SID-2 Medium: 0.585 NDCG in 3.11 s, versus Claude Opus 5.5 (medium) at 0.577 in 25.1 s.
  • SID-2 Swarm, 6 agents: 0.701 NDCG in 17.7 s, versus GPT-6 Astra (max) at 0.704 in 135 s.
  • SID-2 Swarm, 12 agents: 0.720 NDCG, the highest score in the table.
  • Baselines: BM25 scores 0.260 and vector search plus reranker scores 0.287.

Swarm leads on 11 of 19 corpora. On WixQA, a single high-effort agent matches Astra at about 232x lower cost. At its lowest effort, SID-2 beats 4 parallel SID-1 runs with 13x fewer FLOPs.

Source link

Leave a Reply

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