RAG Engine Cross Corpus Retrieval is a document search feature of the Gemini Enterprise Agent Platform, released by Google Research and Google Cloud on June 5, 2026. It is designed to formulate search plans and aggregate relevant context when the evidence required for a single question is scattered across multiple independent RAG corpora. While conventional Retrieval-Augmented Generation (RAG) directly links one question to one search, this feature decomposes complex questions into multiple search plans and sub-queries. It then selectively explores corpora suitable for each query, evaluates whether the acquired context is sufficient for the answer, and repeats the search if necessary. This approach is akin to a librarian not only reading the first book found on one shelf but also sequentially checking the journal room, patent room, and clinical data room to assemble evidence for an answer. Existing RAG pipelines rely heavily on the quality of the initial search results. When questions span multiple domains or required facts are distributed across different storage systems, single-corpus searches may return only partial evidence or miss highly relevant documents. The approach of merging all documents into a single massive index creates operational burdens in environments where corpus-specific access policies, update cycles, and document characteristics must be distinguished. The core differentiator of this feature is that it does not stop at passing the initial search results directly to the generation model; instead, it applies an Agentic RAG loop structure comprising search planning, corpus selection, context sufficiency evaluation, and additional search. Consequently, the process includes determining whether the current evidence alone can address the entire question before generating the answer. Life science researchers can maintain materials with different purposes and update cycles—such as research papers, clinical trial documents, and regulatory guidelines—as separate RAG corpora while processing integrated questions. For example, when asking about the mechanism of action of a specific biomarker, related clinical trial results, and companion diagnostic requirements from regulatory agencies, the system can be configured to execute sub-queries against each data group and compare the scope of the collected evidence. In the evaluation phase, search results from the three corpora can be recorded using metrics such as Recall@10, number of evidence documents, and citation inclusion rate. An experimental design is also possible where, if evidence from a specific data group is missing, only that corpus is re-searched. This is suitable for research reviews where traceability of evidence and detection of omissions per data group are more critical than simple document summarization. Another use case involves candidate compound review and safety investigations. Researchers can construct internal experiment reports, public papers, and toxicity/safety data as independent corpora and submit questions requiring multi-stage judgment, such as whether the target evidence for a candidate compound contradicts reported adverse effects. Search results can be post-processed into evidence tables per compound, document sources, and gaps requiring additional review, allowing personnel to check citation coverage per corpus before making final judgments. However, since supported APIs, corpus connection limits, search iteration control parameters, data security requirements, and regional conditions are not included in this input, official documentation must be consulted prior to operational implementation.

RAG Engine Cross Corpus Retrieval is a document search feature of the Gemini Enterprise Agent Platform, released by Google Research and Google Cloud on June 5, 2026. It is designed to formulate a search plan and aggregate relevant context from multiple independent RAG corpora when the evidence required for a single question is distributed across them. While conventional Retrieval-Augmented Generation (RAG) directly links one query to one search, this feature handles complex questions by performing multiple searches.