To navigate vast amounts of data, SID-2 bets on collective intelligence
Built to explore large data collections, SID-2 combines adaptive search, coordinated agents, and precision controls, with gains evaluated internally.
The expected output from SID-2 is a ranked list of documents useful for answering a question. Available through an API, this search model explores data using tools supplied by the application: it issues queries, reads results, and decides whether to continue or stop. The documentation specifies that users retain their existing search engines and must connect them to the model. Integration does not require reindexing the data; only content returned by the tools is exposed to the system.
The second generation includes a single agent, SID-2, and a collective version, SID-2 Swarm. With the single agent, observations accumulate in one context window, and successive steps add latency. With the swarm, multiple agents explore different paths in parallel, using separate contexts and sharing their findings. Each agent can create additional agents within the task’s limit, request further searches, or pass along relevant references. The documents they identify are then combined into a shared ranking.
The team trained this cooperation through reinforcement learning, with agents working on the same search sharing a reward. Its hypothesis is that document retrieval is particularly suited to this division of labor: checking the relevance of a document already found is often easier than discovering it. Having small models explore more paths could therefore be more effective than increasing the size of a single model. In the technical report, coordinated swarms show greater improvements at comparable compute budgets than simply combining independent searches.
The comparisons cover 506 questions across 19 corpora, including 207 questions from public benchmarks and 299 produced by the team’s query-generation system. At medium effort, the single agent achieves an NDCG score of 0.58 with an estimated inference time of 3.1 seconds, compared with approximately 25 seconds for Claude Opus 5.5 at comparable quality. With six agents, the swarm reaches approximately 0.70 in 18 seconds, a level GPT-6 Astra at maximum effort achieves in 135 seconds. NDCG measures document relevance and ranking position. These times are calculated estimates rather than end-to-end measurements from a production application.
Results vary considerably across collections. The twelve-agent configuration achieves the highest aggregate score and leads on Enron emails and EnterpriseRAG-Bench. The GPT-6 Astra model retains an advantage on California law and some financial corpora, while the swarm outperforms the other systems on PubMed and arXiv. On WixQA’s customer support questions, using multiple agents provides little additional benefit. The evaluation also reveals that some competing models struggle to finish tasks or return the requested results, which affects their scores.
Search behavior remains configurable. Three effort levels control the trade-off between time spent searching and the quality sought. A feature called Metric Steering also directs the model toward broader coverage, stricter selection, or a specified number of results: finding every contract that meets a criterion requires a different strategy from suggesting a few references for a support response. These controls were incorporated into training. The team first rewards retrieval quality, then introduces a latency penalty; when no relevant document is found, finishing faster does not earn an additional reward.