Cloudflare, Weaviate and the RAG debate

Retrieval-augmented generation, or RAG, is the part of an artificial intelligence system that fetches source material before the model answers. This week, Cloudflare and Weaviate used that plumbing to argue over a basic question: when do you need vectors at all? (developers.cloudflare.com) (weaviate.io) Cloudflare said on April 16 that its AI Search service now runs hybrid retrieval, combining vector search for similar meaning with BM25 keyword search for exact terms in one fused result list. The company also added relevance boosting, built-in storage for new instances, and a namespace-level Workers binding that can search across multiple instances in one call. (blog.cloudflare.com) (developers.cloudflare.com 1) (developers.cloudflare.com 2) Cloudflare’s new indexing flow can now take files directly through an Items application programming interface, instead of requiring a separate object-store step in some setups. Its docs say AI Search ingests uploaded files, converts supported formats to Markdown, chunks them, embeds them, and stores them in its managed index. (developers.cloudflare.com 1) (developers.cloudflare.com 2) That product update landed in the middle of a retrieval argument that has been building for months in legal and financial software. For highly structured documents like contracts, invoices, and filings, some builders are leaning harder on filters, fields, and exact term matching instead of relying on semantic similarity alone. (weaviate.io) (docs.weaviate.io) Weaviate’s pitch is not “no search,” but more database-style search directed by an agent that understands schema. Its Query Agent documentation says the service can translate a natural-language prompt into searches, filters, aggregations, and cross-collection routing inside a Weaviate Cloud cluster. (docs.weaviate.io) (weaviate.io) That matters most when the question includes a hard constraint like year, counterparty, or contract type. Weaviate’s legal RAG example says its own finance team used Query Agent to navigate internal contracts, and the product page says the agent applies dynamic filters and query optimization before returning cited results. (weaviate.io 1) (weaviate.io 2) Cloudflare’s update points the other way: keep vectors, but stop treating them as the only retrieval method. Its April 16 changelog says hybrid mode runs vector and BM25 in parallel, then fuses the rankings, a setup aimed at questions that need both synonym matching and exact language. (developers.cloudflare.com) Weaviate also supports BM25 keyword search alongside vector search in its database, so the dispute is less about whether keywords matter than where precision should live. In Weaviate’s framing, the agent can reason over the database structure itself; in Cloudflare’s framing, the managed search layer handles mixed retrieval and feeds agents from there. (docs.weaviate.io) (docs.weaviate.io) (blog.cloudflare.com) Cloudflare is also renaming part of the stack as it goes. The company’s April 16 post describes AI Search as “formerly AutoRAG,” a service Cloudflare had introduced in open beta on April 7, 2025 as a managed retrieval-augmented generation pipeline. (blog.cloudflare.com) (blog.cloudflare.com) The practical split is getting clearer: broad, messy corpora still reward semantic retrieval, while contracts and financial records often reward exact fields, exact dates, and exact clauses. The current product launches suggest vendors are converging on a mixed model, even as they keep arguing over which part should lead. (developers.cloudflare.com) (weaviate.io)