Ask the whole corpus, not a lucky excerpt

Two modes, one honest decision rule

Full-window Q&A: when the set fits in 1M tokens, load it all and ask. The model reasons across every document simultaneously — the only reliable way to catch contradictions, spot omissions, and reconstruct timelines that span files.

Retrieval Q&A: when the corpus exceeds any window or changes daily, Bee's retrieval pipeline indexes it and grounds each answer in cited passages. Same endpoint, different tool for a different corpus shape.

Questions that only whole-corpus reading answers

• →Which obligations in this contract stack conflict with each other, and where?
• →Across a quarter of meeting transcripts, when did this decision actually change — and who changed it?
• →Do these policy documents anywhere contradict the procedure manual?
• →What did every incident report this year have in common?

Answers you can check

Retrieval answers carry cite-backs to source passages, so verification is a lookup, not a re-read. For full-window requests, token usage is reported exactly as consumed, and model aliases resolve to dated snapshots — the answer you got is attributable to a specific release over specific input.

Capacity and where to run it

The 1M-token long-context mode is served on Bee Comb and Buzz; Bee Swarm runs a 1M window natively. Retrieval capacity scales by plan — up to 1,000 documents on Comb, 25,000 on Hive, and 100,000 on Swarm — with shared team indexes from Buzz upward.

Customer data is not used to train the Bee base model. Prompts, documents, and outputs stay yours; processing is limited to serving, securing, and operating the service.