LoginSign Up
← Back to all sparks
W

Weaviate

DEVOPS
Velocity5.0

Open-source vector database for AI applications and semantic search

weaviate.io ↗

Weaviate is tuning the agent, not the database — Search Mode now has effort tiers.

vector-databasequery-agentretrievalmcpobservability
◆Current state
Since the 1.38 release took the HFresh disk-based index and the built-in MCP server to general availability, the visible work has moved up the stack. The Query Agent's Search Mode gained medium, high and ultrahigh effort tiers, letting callers trade latency for thoroughness on a per-query basis. Query profiling returns per-stage, per-shard timing so a slow query can be diagnosed rather than guessed at. The remaining entries are a two-part Foundry series arguing that folders, tags and keyword search break down in creative workflows, plus an ingestion guide covering server-side batching and the blobHash type.
◆Where it's heading
The database work is largely banked; what is being iterated now is the agent layer sitting on top of it, and effort tiers are the clearest sign Weaviate treats retrieval quality as a dial rather than a fixed property. Query profiling serves the same shift — once agents issue the queries, humans need a way to see where time went. The Foundry posts are positioning for that same audience: retrieval as the replacement for manual file organisation.
◆Prediction
Expect the effort tiers and profiling to converge — surfacing the cost of a chosen tier — and the Boost API and nested object filtering previews from 1.38 to reach general availability.

◆Recent moves

  1. 6d ago

    Building Foundry Part 2: Where creative workflows break

    Part two of the Foundry series, on why folders, tags and keyword search fail in creative workflows. Positioning content for the retrieval argument, not a product change.

    View source ↗
  2. 8d ago

    Scaling Test-Time Compute in Search Mode

    Adds medium, high and ultrahigh effort tiers to the Query Agent's Search Mode, making retrieval thoroughness a per-query choice rather than a fixed behaviour. It extends the agent layer that 1.38's MCP server made addressable.

    View source ↗
  3. 20d ago

    Building Foundry: AI isn’t replacing creativity, it’s removing friction

    The first Foundry post, arguing AI removes friction from creative workflows rather than replacing creatives. Narrative marketing ahead of the part-two retrieval argument.

    View source ↗
  4. 29d ago

    Query Profiling: See Where a Slow Query Spends Its Time

    Query profiling returns per-stage, per-shard timing breakdowns so slow queries can be diagnosed directly. Necessary observability once agents rather than humans are issuing the queries.

    View source ↗
  5. 1mo ago

    Weaviate 1.38 Release

    ⚡ SPARK

    The 1.38 release takes the HFresh disk-based vector index and the built-in MCP server to general availability, rebuilds async replication onto a single cluster-wide scheduler, and previews a Boost API and nested object filtering. It is the release that made the agent-runtime direction concrete.

    View source ↗
  6. 2mo ago

    Import & Vectorize Data with Weaviate at Scale

    A guide to server-side batching, retries, the blobHash type and multimodal ingestion. Documentation of existing capability rather than new capability.

    View source ↗