Skip to main content
A futuristic digital interface displays "SID-2 Agentic Search Model" with a graph showing 0.720 NDCG Swarm Mode.

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

SID-2 Agentic Search Beats GPT-6 Astra 0.720 NDCG

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

• 2 min read

SID AI put a number on its new retrieval model this week: 0.720 NDCG, posted by SID-2 Swarm running 12 agents in parallel. That score edges out GPT-6 Astra's 0.704, according to figures SID AI released alongside the launch of SID-2 and its swarm variant. The company is pitching SID-2 as a replacement for the standard RAG pipeline, the fixed sequence of rewrite, embed, search and rerank that most retrieval systems still run in lockstep.

Instead, SID-2 works as an agent: it searches, reads what comes back, decides whether that's enough, and searches again if it isn't. Developers wire two tools into their own backend, one for keyword search and one for vector search, and the model closes out each session by calling report_helpful_ids with its ranked list of document IDs.

SID-2 runs with a 262,000-token context window according to SID AI's own report, though the company's model page lists 256k. It's available now through a hosted API at platform.sid.ai, with reference paths for Baseten and SageMaker. Latency numbers in SID AI's materials assume a single B200 GPU handling eight concurrent queries, a detail that matters once you start comparing SID-2 against the models it's being benchmarked against.

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.

Why this matters

For teams building retrieval systems, SID-2 is worth watching but not worth switching to blind. A 506-question benchmark across 19 corpora, including HotpotQA, Enron emails, SEC filings, and PubMed, gives some texture to the 0.720 NDCG figure, but it's SID's own benchmark on SID's own model. The proposer-solver self-play loop against a frozen SID-1 solver is a genuinely interesting training approach, and a 6x throughput gain over SLIME's default setup is a concrete, checkable claim, assuming someone replicates it outside SID's walls.

The practical details matter more than the headline number: a 262k-token context window (SID's own site says 256k, which is its own small red flag), hosted access through platform.sid.ai, and deployment paths via Baseten and SageMaker. Latency numbers are pegged to a single B200 GPU, so anyone benchmarking this against their own RAG pipeline should replicate that hardware baseline before trusting the comparison. Agentic search that decides what to query next instead of following a fixed pipeline is a real architectural shift. Whether it holds up outside curated benchmarks is the next thing to check.

Common Questions Answered

How does SID-2's NDCG score compare to GPT-6 Astra?

SID-2 Swarm achieved a 0.720 NDCG score running 12 agents in parallel, which edges out GPT-6 Astra's 0.704 NDCG score. This performance comparison was released by SID AI alongside the launch of SID-2, demonstrating improved retrieval capabilities over the competing model.

What is the key difference between SID-2 and traditional RAG pipelines?

SID-2 replaces the fixed sequence of rewrite, embed, search, and rerank operations that standard RAG pipelines run in lockstep with an agentic approach that dynamically decides what to search next. Rather than following a predetermined workflow, SID-2 acts as an agent that searches, reads results, and returns ranked document IDs based on intelligent decision-making.

What datasets were used to benchmark SID-2's performance?

SID-2 was evaluated on a 506-question benchmark across 19 different corpora, including HotpotQA, Enron emails, SEC filings, and PubMed. This diverse dataset selection provides texture to the 0.720 NDCG performance metric across multiple document types and domains.

What training approach does SID-2 use according to the article?

SID-2 employs a proposer-solver self-play loop against a frozen SID-1 solver as its training approach. The article describes this as a genuinely interesting training methodology that contributes to the model's improved performance.

What throughput improvement does SID-2 provide over SLIME?

SID-2 delivers a 6x throughput gain over SLIME's default setup, which represents a concrete and checkable performance improvement for teams building retrieval systems. This significant throughput advantage makes SID-2 an attractive option for organizations seeking faster retrieval performance.

LIVE09:42SID AI Releases SID-2 Agentic Search Model With 0.720 NDCG Swarm Mode