Build a private research surface that shows where an answer came from, how the evidence was retrieved, and whether the final response stays grounded in the sources you control.
The August revision adds a review checkpoint for transfers outside approved counterparties and records the reviewer identity in the audit export.1 Existing internal transfers retain the prior threshold.2
Private knowledge is fragmented across websites, documents, and application boundaries. Even after it is indexed, a plausible answer can outrun its evidence. Cosift keeps the retrieval path visible and supports evaluation alongside generation, so teams can test source coverage, ranking, and answer grounding rather than relying on fluency alone.
Choose the sources and retain control over the index, administration surface, and model endpoints.
Use BM25, optional embeddings, reranking, expansion, HyDE, and diversity controls for the query at hand.
Query, answer, and multi-step research endpoints return source context instead of hiding the retrieval trace.
Cosift keeps collection, retrieval, generation, and evaluation as separate observable stages. Teams can improve the stage that failed instead of treating search quality as model intuition.
Crawl the approved public or private knowledge surfaces.
Start with BM25 or add dense retrieval and reranking where they improve measured quality.
The agent receives the response together with source references.
Built-in retrieval and answer checks make quality a repeatable engineering concern.
Share the shape of the knowledge base and one question that matters. We will outline the crawl, retrieval, citation, and evaluation path worth testing.