Vespa! Cloud Infrastructure for Search Applications!

Vespa is one of the most mature and performant vector-aware search engines in production today — originally developed at Yahoo and now powering AI search at scale for global companies. It natively combines text and vector retrieval, supports streaming writes, online re-ranking, and large-scale low-latency inference in a single platform.

• Trusted by Yahoo, Perplexity, Vinted, FARFETCH, Onyx (Danswer Cloud), Splore.
• Combines ANN (approximate nearest neighbor) retrieval with BM25 keyword ranking, dynamic filtering, and tensor computation for real-time personalization.
• Designed for multi-node scaling — horizontal sharding, coverage tracking, and replica rebalancing are built into the serving layer.
• Fully open-source and Kubernetes-ready, with deep documentation at docs.vespa.ai.

schema arxiv2 {
  document arxiv2 {
    field arxiv_id type string { indexing: attribute | summary; fast-search: true }
    field title type string { indexing: index | summary; match: bm25 }
    field abstract type string { indexing: index | summary; match: bm25 }
    field categories type array<string> { indexing: attribute | summary; fast-search: true }
    field family type string { indexing: attribute | summary; fast-search: true }

    field title_Embedding_1 type tensor<bfloat16>(d0[384]) {
      indexing: attribute
      attribute: fast-search
      distance-metric: angular
    }
    field Summarized_Abstract_Embedding_1 type tensor<bfloat16>(d0[384]) {
      indexing: attribute
      attribute: fast-search
      distance-metric: angular
    }
  }

  fieldset default { fields: title, abstract }

  rank-profile keyword inherits default {
    first-phase { expression: bm25(title) + bm25(abstract) }
  }

  rank-profile ann_title_1 inherits default {
    inputs { query(query_embedding) tensor<bfloat16>(d0[384]) }
    first-phase { expression: closeness(title_Embedding_1, query_embedding) }
  }

  rank-profile ann_summary_1 inherits default {
    inputs { query(query_embedding) tensor<bfloat16>(d0[384]) }
    first-phase { expression: closeness(Summarized_Abstract_Embedding_1, query_embedding) }
  }

  rank-profile ann_multi_1 inherits default {
    inputs {
      query(query_embedding) tensor<bfloat16>(d0[384])
      query(w_title) double
      query(w_abs) double
    }
    first-phase {
      expression: query(w_title) * closeness(title_Embedding_1, query_embedding)
                + query(w_abs)   * closeness(Summarized_Abstract_Embedding_1, query_embedding)
    }
  }

  # Fine-tune style profile example
  rank-profile ann_fine_tune_1 inherits default {
    inputs { query(query_embedding) tensor<bfloat16>(d0[384]) }
    first-phase {
      expression: closeness(title_Embedding_1, query_embedding) +
                  pow(max(0, closeness(Summarized_Abstract_Embedding_1, query_embedding)), attribute(weight1)) * attribute(weight2)
    }
  }
}

• Horizontal: increase replicas for vespa-query and vespa-feed to raise QPS/ingest.
• Vertical: grow container resources and content memory if paging/latency spikes.
• Autobalancing: content cluster rebalances buckets after scale; brief <100% coverage is normal.