SAG is a document search tool released by Zleap AI on July 14, 2026. It extracts complete events and the entities involved within sentences, then constructs both relational and vector indexes. While conventional Retrieval-Augmented Generation (RAG) focuses on splitting documents into fixed-length chunks and finding semantically similar chunks, SAG adopts an approach that preserves event-level units corresponding to who did what, when, and how. Upon receiving a query, instead of traversing only pre-fixed graphs, SAG creates dynamic hyperedges using SQL joins and performs multi-hop searches connecting multiple events and entities. This is closer to assembling a temporary investigation board linking people, events, and documents tailored to the question, rather than merely finding pages with similar sentences in a library. Existing Dense RAG excels at quickly retrieving documents with similar expressions but may miss intermediate relationships when questions require connecting facts scattered across multiple documents to form an answer. Conversely, GraphRAG approaches are strong in relationship exploration but incur costs for pre-defining and maintaining graph schemas and edges. The differentiator of SAG is that it does not simply merge the results of Dense RAG and GraphRAG; instead, it indexes event–entity structures and vector representations during the collection phase and dynamically generates relationships required by actual queries during the search phase. Consequently, even with the same document set, it can construct different connection paths depending on the query, aiming for a structure suitable for tracking event contexts or common participants rather than simple similar-phrase searches within single documents. Life science researchers can build an event-centric search foundation in SAG by inputting materials where the same genes, drugs, and diseases appear repeatedly in different contexts, such as research papers, clinical trial public documents, and lab records. For example, by linking entities such as experiments inhibiting a specific target protein, observed phenotypes, and cell lines used, researchers can explore questions that require traversing multiple events, such as "When this target was inhibited, inflammatory markers decreased, but under what conditions were toxicities reported?" Search results can be passed to summarization models via REST or OpenAI-compatible APIs, or used as the evidence retrieval stage for research assistant agents through MCP integration. However, supported document formats, extraction models, chunking settings, reranking methods, and quantitative performance cannot be verified solely from the provided Discovery information, so official documentation verification is necessary before actual research application. Another use case involves connecting reagents, batches, treatment conditions, and observation results in long-term accumulated experimental records into event–entity forms. Researchers can choose local desktop or Docker deployment and track conditions scattered across different lab notebooks based on SQL join paths generated at query time. In literature monitoring tasks, events extracted from new papers can be added to existing indexes, followed by multi-hop queries across drug–target–disease–adverse event relationships to narrow down candidates for subsequent review. However, local processing scope, external model invocation status, data retention policies, and API authentication methods are not included in the input information; therefore, deployment settings and security boundaries must be verified separately before injecting sensitive clinical or genomic data.

SAG is a document search tool released by Zleap AI on July 14, 2026, which extracts complete events and the entities involved in each event from sentences, then constructs both relational and vector indexes. While conventional Retrieval-Augmented Generation (RAG) focuses on splitting documents into fixed-length chunks and retrieving semantically similar chunks, SAG adopts an approach that preserves event-level units corresponding to who did what, when, and how. When a query arrives, instead of traversing only a pre-fixed graph, it dynamically performs SQL joins

🔬 Literature-Based Drug–Target–Phenotype Tracing: Index drugs, target proteins, cell lines, and observed results from papers as event–entity pairs, then perform multi-hop queries using SQL joins. Supported file formats, extraction parameters, and search performance metrics must be confirmed via official documentation.