New Search Effort Tiers and Weaviate 1.39

Hello Weaviate Community! 🤗

This week, we release Weaviate 1.39, bringing the Boost API and MMR diversity selection to GA. Query Agent now offers medium, high, and ultrahigh search-effort tiers, while our Explorer in the console lets you run keyword, semantic, and hybrid search without requiring hand-written GraphQL.

On the podcast, we talk with Alex Zhang about Recursive Language Models and Mathew Jacob about why cross-encoder rerankers can lose recall as candidate sets grow.

Latest AI & tech insights

Explore our recent Weaviate content:

Read

• 🚀 Weaviate v1.39 Release: The Boost API and MMR diversity selection now GA, with 4-bit rotational quantization in preview. The release also adds an experimental Search REST API and automatic HNSW snapshots. Read the blog

• ✍️ Scaling Test-Time Compute in Search Mode: See how the new effort parameter balances retrieval accuracy, latency, and cost across eight reasoning-intensive and domain-specific benchmarks. Read the blog

• 🔖 Building Foundry Part 2: Where creative workflows break: Folders, tags, and keyword search lose effectiveness as creative projects grow and naming conventions drift. Learn why retrieval needs to understand intent even when metadata is incomplete. Read the blog

Watch

• 🎙 Recursive Language Models with Alex Zhang: Alex Zhang explains how Recursive Language Models treat prompts as program variables, letting models manipulate context and spawn recursive calls over smaller pieces. The conversation also covers PrimeAgent's persistent IPython REPL, long-horizon benchmarks, and why the team moved from prompt tuning to reinforcement learning on Qwen3. Watch the full podcast

• 🎙 Drowning in Documents with Mathew Jacob: Mathew Jacob explains why cross-encoder rerankers can lose recall when reranking beyond roughly 100 documents, including experiments where BM25 beat state-of-the-art cross-encoders over 10,000 candidates. Watch the full podcast

Product highlights

Check out our latest product updates.

• 🔎 Search Mode effort tiers: Choose medium, high, or ultrahigh effort per request to control how much work Query Agent puts into query writing and reranking. On BRIGHT Biology, ultrahigh effort reached 57.5 nDCG@10 compared with 13.0 for Hybrid Search. Explore the benchmarks

• ☁️ Keyword, semantic and hybrid search in Explorer: Explorer now runs keyword, semantic and hybrid search plus aggregations, with no hand-written GraphQL. Filter in any mode, tune the keyword-to-vector balance on hybrid, and hit Find similar on any result to pull up its nearest neighbors. Try it now

• 🚀 Weaviate 1.39 Highlights:

  • Boost API: Promote or demote results at query time without filtering them out.

  • MMR diversity selection: Bring more variety to hybrid and vector result sets.

  • 4-bit rotational quantization (Preview): Store a 1536-dimensional vector in 784 bytes, 7.84× smaller than raw float32.

  • Search REST API (Experimental): Run search over plain HTTP with JSON requests and responses.

  • Automatic HNSW snapshots: Compact index state and trim covered commit logs in the background for less disk use and steadier restarts.

  Read the release blog for the full details.

Hungry for more?

Have a question or want to connect? Join our Weaviate Forum to engage in community conversations.

Follow our Blog, LinkedIn, Twitter, and YouTube for more content updates.

See you in two weeks,

Prajjwal