Overview

SID-1 is an agentic retrieval model. It is trained to find and rank information across knowledge bases, finance, legal, science, and email. It can deliver deep research recall at close-to-static pipeline speeds.

A horizontal bar chart comparing seven retrieval approaches across two metrics: Retrieval Performance (Recall, where higher is better) on the left and Latency (Seconds per Question, where lower is better) on the right. The rows from top to bottom are SID-1 (4x), GPT-5.1 (high), SID-1, Gemini 3 Pro, Sonnet 4.5, Reranker @10, and Vector only @10. The two SID-1 rows are drawn as solid, bold bars and the other approaches are lighter, so the SID-1 results stand out. Recall values: SID-1 (4x) 0.84, GPT-5.1 (high) 0.78, SID-1 0.77, Gemini 3 Pro 0.66, Sonnet 4.5 0.64, Reranker @10 0.45, Vector only @10 0.44. Latency values: SID-1 (4x) 5.5s, GPT-5.1 (high) 131s, SID-1 5.5s, Gemini 3 Pro 156s, Sonnet 4.5 35s, Reranker @10 0.78s, Vector only @10 0.15s. The point of the chart is that SID-1 reaches recall close to the strongest frontier LLM setups while running more than twenty times faster than those frontier LLM loops, close to static pipeline speeds.

Static vs agentic retrieval

A static retrieval pipeline runs a fixed sequence of steps and returns whatever it found, even when that is wrong. It cannot react to what it sees. In practice, static pipelines accumulate a long tail of failures on tasks like multi-step or multi-hop questions.

A static retrieval pipeline diagram. The flow moves left to right from a question box to an optional LLM rewrite box, then to a search backend box, then to an optional reranker box, and finally to a results box. The diagram represents a fixed design-time retrieval sequence: each query follows the same predetermined stages, and the pipeline does not loop back to inspect retrieved content or choose new searches based on intermediate results.

SID-1 works the way a person does: it searches, reads the results, adapts, and searches again until it has enough. SID-1 can react to how difficult a question is. If it is easy and the results are found quickly, it can choose to end early. If the question is hard, it will exert more effort until it finds the right result or exhausts its turns. This pattern generalizes to arbitrary questions: If a human could find the information, then an agentic retrieval model like SID-1 can too.

A SID-1 retrieval pipeline diagram. A question flows into SID-1, and SID-1 returns ranked results. Between SID-1 and the search backend, three parallel downward arrows show repeated tool calls such as BM25, ANN, and other search operations over multiple turns. Three parallel upward arrows show the search backend returning content and metadata to SID-1. The diagram emphasizes that SID-1 iterates with the search backend before producing the final ranked result list.

This iterative approach comes at the cost of latency. When implemented with frontier LLMs, this loop can take ~30-150s, which is too slow for many user interfaces. Because SID-1 is a model specialized for this task, we were able to make the model ~20x faster: 1.5s (easy) to 8s (very hard) per question. This is still slower than many static pipelines (sub-1s), but makes agentic search practical for many more applications.

SID-1 is not a search backend

You own your data and the code around the model. You implement the search tools that run against your backend (your index, your document IDs), format the results, and run the loop. Using SID-1 doesn't require you to reindex your data or store it somewhere else. It is designed to work with any search backend or database you already have.

SID-1 owns the decisions inside the loop: which tool to call, how to phrase each query, which documents to keep, and when to stop. It never sees your corpus directly; it only sees the results your tools return.

Content

Static vs agentic retrievalSID-1 is not a search backend