AI Engineer
← All organizations

AI search and retrieval models

SID.ai

SID.ai trains AI models to find the documents other AI systems need to answer questions. Its SID-1 agentic retrieval model accepts natural-language queries, searches customer-selected document collections and returns ranked results. Developers can integrate it into existing search or retrieval-augmented generation systems, or use it as a search subagent for a larger language model. Returning documents rather than synthesized answers lets the downstream system handle answer generation while SID-1 handles retrieval. SID describes access through an API, AWS Bedrock and self-hosted deployment, with a gradual rollout and a public waitlist.

Founded in 2023 by Maximilian-David Rumpf and Lotte Seifert, SID is led by Rumpf as CEO. Rumpf previously researched AI accelerator design at ETH Zürich; Seifert holds degrees in computer science from ETH Zürich and economics from the University of St. Gallen. The company has a team in San Francisco and Zürich and lists Y Combinator, Canaan, Rebel and General Catalyst as backers.

SID’s retrieval research applies multi-turn reinforcement learning to Qwen3-14B, training the model to search, read results and adjust its approach to each question. Its reward measures both whether it finds relevant documents and how well it orders them, so downstream systems can take the highest-ranked results when context space is limited. SID deliberately favors finding an extra document over omitting a crucial one, accepting some loss of precision. Parallel tool calls reduce round trips to the search backend, while initial searches return short excerpts and a separate read tool lets the model selectively open full documents. This gives it room to inspect more candidates without filling its context with every document’s full text.

SID’s service terms describe self-serve and enterprise access, with pricing options that may include subscription minimums and usage-based charges. Customers retain ownership of their inputs and outputs, including submitted documents and returned rankings. SID commits not to use that customer content to train, fine-tune, distill or otherwise improve AI models; the commitment survives termination of the terms.

Explore the recordings

The supplied SID.ai archive contains one recording. It offers a path from retrieval tradeoffs to reinforcement learning for specialized search; the guide below organizes that recording by viewing goal.

Start with the retrieval problem

Watch Where RL Will Take Search — Maximilian-David Rumpf, SID.ai for the argument behind search sub-agents. Rumpf contrasts fixed vector-and-reranker pipelines with agentic search, then describes a learned policy that searches, inspects documents, adjusts filters and retries. This is a useful starting point for understanding adaptive retrieval and the proposed benefit of keeping intermediate search results outside the main agent’s context.

Follow the training and evaluation argument

Return to the same recording to examine document-level rewards, compute scaling and latency-aware training. Rumpf reports approximately 20-fold faster search and 100-fold lower cost for SID-1 than a frontier model on the presented task, while acknowledging that classical pipelines remain faster. Treat these figures as recorded benchmark claims, not independently verified comparisons or statements about current performance. His proposed applications in voice, e-commerce and private enterprise knowledge are forecasts in the recording, not evidence of today’s product availability.

1 talk

Newest first

1 speaker at AIE

Affiliations reflect their AIE appearances, not necessarily current employment.

Maximilian-David Rumpf
Company sources · checked 2026-09-18