Weaviate demo: 24‑hour legal AI
Weaviate published a 24‑hour production case study building a legal AI using a 'Query Agent' that combined multivector embeddings, schema routing and agentic search to produce precise, cited results. The demo emphasises schema‑aware retrieval, routing across vector spaces, and agent coordination for domain‑heavy workflows. It serves as a concrete pattern for adapting RAG and agents to regulated, citation‑sensitive SaaS documents. (x.com)
Most legal question-answering systems fail on the boring part, not the clever part. Ask for a notice clause from a 2024 service agreement, and a basic retrieval system can still hand you a similar clause from 2022 because it matched the wording instead of the date filter. (weaviate.io) Retrieval augmented generation is the standard pattern here: search documents first, then let a language model write the answer from the retrieved text. That works for simple help-center questions, but legal documents add jurisdiction, contract type, date ranges, and clause-level precision that a one-shot search often misses. (weaviate.io) Weaviate’s demo was a legal assistant built in 36 hours for contract questions, using the company’s Query Agent on top of a public demo dataset that replaced its internal finance contracts. The point was not a chatbot skin; the point was showing how the search logic itself could be delegated to an agent that plans the query before it searches. (weaviate.io) A Query Agent is Weaviate’s hosted service that takes a natural-language question and decides the database operations needed to answer it. The service is available on Weaviate Cloud, and its documentation says it can inspect collections, run searches, use aggregations, and return answers with intermediate results. (docs.weaviate.io) That “inspect collections” step is the quiet trick. Instead of treating every document pile like one giant folder, the agent reads collection and property descriptions first, so it knows which fields can be filtered and which collections are even relevant to the question. (docs.weaviate.io) Weaviate describes this as treating the database like a toolbox instead of a filing cabinet. In its March 5, 2025 launch post, the company said the Query Agent can route a question across multiple collections and decide whether the job is semantic search, aggregation, or both. (weaviate.io) The retrieval side also changed. Weaviate added multi-vector embeddings in version 1.29, which means one document can be represented by many vectors instead of one compressed summary vector. (docs.weaviate.io) That matters for contracts because a single sentence can carry the whole risk. Weaviate’s docs explain that multi-vector search uses “late interaction,” which compares smaller parts of text instead of forcing the whole document through one similarity score, so a clause about termination or indemnity is less likely to disappear inside a long agreement. (docs.weaviate.io) The legal demo also used sub-agents for reranking and answer writing. In Weaviate’s case study, the rerank step reordered matches by actual relevance to the user’s question, and the answer step produced a grounded response from that narrower set of passages. (weaviate.io) Weaviate spent February 2026 pushing this from product feature to repeatable recipe. On February 18, 2026, it published Agent Skills, a repository of scripts and cookbooks for coding agents, including schema inspection, data ingestion, and direct Query Agent integration for “Ask” and “Search” modes. (weaviate.io) So the demo is less “we built a lawyer bot over a weekend” than “we turned legal search into a chain of explicit database decisions.” In regulated software, where every answer needs a source and every filter has to survive scrutiny, that is the difference between a fluent demo and a tool someone in finance or legal might actually trust. (weaviate.io)
