September 2026 Summaries

Retrieval quality is presented as a central challenge in AI agent design because agents must first gather accurate, relevant context before language models can produce reliable answers or actions. Failures such as hallucinations, excessive context, and latency often originate in weak search, incorrect database or API calls, poor filtering, or low-ranked relevant sources rather than in the generation model itself. The post argues that systems need retrieval traces and evaluations that record queries, filters, returned results, selected context, and relevance labels to identify whether information was missed during query construction, candidate retrieval, ranking, summarization, or context assembly. It describes a common architecture in which applications fan out to multiple context-building tools and then consolidate evidence for generation, with Vespa positioned as a retrieval layer for large, dynamic, permissioned corpora requiring hybrid keyword and semantic search, metadata filtering, ranking, and controlled summaries. Effective ranking combines signals such as exact term matches, semantic similarity, metadata, freshness, and optional reranking models, helping systems pass fewer but more useful sources to models while reducing token costs and latency.

Four Vespa interns developed a standalone, publicly hosted Model Context Protocol server that enables AI assistants such as Claude and Codex to deploy, inspect, discover, monitor, and troubleshoot Vespa Cloud applications without requiring local installation or an application-specific server. The project emphasizes careful MCP tool design, using intuitive parameters, concise responses, filtering, summaries, and informative errors to reduce agent mistakes and context costs, while adding capabilities such as direct access to Grafana metrics that are not available through the Vespa CLI. The team addressed protocol limitations involving file uploads and long-running deployments with HTTP upload endpoints, configurable blocking behavior, and plans to adopt newer MCP task features. Authentication works for Vespa Cloud’s Control Plane through Auth0, but secure multi-tenant Data Plane access remains unresolved, limiting the prototype’s production readiness. Evaluations comparing MCP-enabled agents with agents using only a terminal and Vespa CLI found similar ultimate task success rates, but MCP agents required fewer deployment attempts, resolved scenarios more cleanly, and made fewer errors, though tool descriptions and discovery added token overhead. The server is undergoing security review and requires further guardrails and user feedback before a potential public release.

Vespa’s September 2026 newsletter announces product enhancements for AI-powered retrieval, ranking, cloud operations, and observability, alongside its first in-person community event, Vespa.ai Live in London. New capabilities include time-bounded approximate nearest-neighbor vector search for meeting latency budgets, a max-cost-factor option that keeps Vespa Cloud clusters scalable when preferred cloud instances are unavailable, and expanded query labeling features that enable separate BM25 and raw-score contributions for subqueries. Additional ranking tools expose array offsets, field-length statistics, query-term document frequencies, and multidimensional structured-data tensors to support more flexible custom ranking models. Vespa Cloud Enclave users can now export metrics and logs directly to OpenTelemetry-compatible observability systems, while new Asia linguistics sample applications demonstrate approaches to Chinese-language search across regional variants. Other updates cover document ID indexing, disk-size resampling, optimized Hamming MaxSim, new videos and research content, and upcoming appearances at events in London, Trondheim, and Nashville.