SID AI launches SID-2 agentic search model with Swarm mode

Executive Briefing
“SID AI has released SID-2, its second-generation agentic search model designed to...”
Key Takeaways
- 1.SID-2 replaces static RAG pipelines by deciding what to search next and returning ranked document IDs.
- 2.SID-2 Swarm using 12 agents achieved an NDCG score of 0.720 on SID's self-reported benchmark, outperforming other tested models.
- 3.Training utilized a two-stage reinforcement learning setup, with the second stage adding a multiplicative latency penalty on positive rewards.
SID AI has released SID-2, its second-generation agentic search model designed to autonomously execute iterative retrieval steps. The system replaces static RAG pipelines with a mechanism that decides its next search path dynamically.
Product Features & Key Highlights
SID-2 replaces static RAG pipelines by deciding what to search next and returning ranked document IDs.
SID-2 Swarm using 12 agents achieved an NDCG score of 0.720 on SID's self-reported benchmark, outperforming other tested models.
Training utilized a two-stage reinforcement learning setup, with the second stage adding a multiplicative latency penalty on positive rewards.
Practical Assessment & Evaluation
Shifting from static RAG pipelines to autonomous agent loops prioritizes document reading depth over raw single-pass LLM reasoning.
The Metric Steering mechanism enables targeted optimization for NDCG, F1, or Recall, enhancing practical enterprise utility.
Pricing, Availability & Next Steps
Specific parameter sizes and model weights have not been publicly disclosed by the developer.
Performance benchmarks rely entirely on self-reported metrics from SID AI without independent third-party evaluations yet.







