Weaviate adds agentic search features
Weaviate released Agentic Search with query agents that inspect schema, route intelligently, and rerank results — the team tested the feature on legal contracts for conversational queries. The capability embeds agentic orchestration directly into the vector store layer, changing how RAG pipelines can offload routing and schema reasoning. (x.com)
Weaviate moved the Query Agent to general availability on September 17, 2025, and the feature is offered as a managed service on Weaviate Cloud rather than as a self-hosted add‑on. (weaviate.io blog; docs.weaviate.io) The Query Agent performs schema inspection, decomposes multi‑step queries, chooses between regular semantic search and aggregations, and constructs dynamic filters and structured queries before executing searches. (weaviate.io blog; docs.weaviate.io) Weaviate demonstrated the Query Agent by building a full‑stack Legal Contract RAG application in 36 hours using the CUAD dataset and internal contract data, citing schema & collection inspection and automated structured query construction in that effort. (weaviate.io blog) Responses from the agent include structured answers with source citations that reference collection names and object IDs, and the service exposes both “ask” (synthesized answers) and “search” (raw object results) modes. (weaviate.io product pages; GitHub weaviate-agent-skills) Public tooling and examples shipped alongside the launch: a Colab “query‑agent‑get‑started” notebook, a weaviate_query_agent_demo GitHub repo, and a weaviate‑agents PyPI package that documents QueryAgent and a Transformation Agent (marked alpha). (colab.research.google.com; github.com/weaviate/weaviate_query_agent_demo; pypi.org/project/weaviate-agents) Multiple third‑party integrations and evaluations show the Query Agent acting as a callable tool in agentic RAG setups—LlamaIndex published a multi‑agent workflow using Weaviate’s Query Agent, and TruLens released an evaluation cookbook for the Query Agent—illustrating patterns where orchestration, collection selection, and reranking are handled inside the vector DB layer. (developers.llamaindex.ai; trulens.org)
