AgentsApplications 🇺🇸 24.07.2026 00:02

Agentic Retrieval for Amazon Bedrock Managed Knowledge Base

Amazon Web ServicesAmazon Web Services
Amazon Bedrock Managed Knowledge Bases introduces agentic retrieval, a new mode that decomposes multi-part, comparative, and exploratory questions into sub-queries, retrieves iteratively, and can generate a grounded response in one API call. It shows up to 37% absolute recall improvement on multi-hop benchmarks.
Amazon Bedrock Managed Knowledge Bases now offers agentic retrieval via the AgenticRetrieveStream API, which uses a foundation model to plan and iterate over retrieval steps instead of a single similarity search. It handles multi-intent, comparative, and cross-source queries by decomposing the question, retrieving per sub-query, evaluating sufficiency, and optionally generating a grounded answer. The API returns a stream of trace events (speculative retrieval, planning, retrieval steps, and final result), supporting up to five knowledge bases with natural-language descriptions for routing. Benchmarking on MuSiQue shows up to 37% absolute recall improvement on 4-hop questions. Pricing includes $4 per 1000 calls with managed model plus $1 per 1000 underlying Retrieve calls. Users can tune maxAgentIteration (default 3 for single KB, 4-5 for multi-KB) and choose between service-managed or custom foundation models.
Сокращения
KB = Knowledge Base — База знаний
API = Application Programming Interface — Программный интерфейс приложения
FM = Foundation Model — Фундаментальная модель
Source: AWS ML blog — original
Our earlier posts on this topic ↓
Fresh news