What Happened
Vespa added fast-search map fields for attributes stored as key-value pairs. Instead of finding documents with a matching key and value and then checking whether they occur at the same array position, Vespa indexes each pair as a combined term for direct lookup. In a benchmark on a Graviton 4 node, this improved performance by 100× or more; that result should not be assumed for other workloads. The feature is available in Vespa 8.773.18 and supports map lookups through mapMatch() or shorthand syntax. Vespa also added string range queries in 8.755.13, enabling an encoded approach to numeric ranges in these fields [1].
Cohere released Embed 5 Pro and the lighter Embed 5 Fast. Both accept text and images, support more than 100 languages, and share an embedding space [2]. The available details do not establish a performance or cost comparison against other embedding models.
Why It Matters to Businesses
RAG quality depends on retrieving the right authorized material, not just choosing a capable language model. For catalogs and other records with variable attributes, a slow structured-filter stage can delay retrieval even when vector search is fast. Direct key-value indexing offers a way to improve that stage where the query and data shape fit [1].
Text-and-image embeddings create an option for searching documents that contain both prose and visual content, including PDFs [2]. Businesses should still test whether a shared embedding space retrieves the specific pages, figures, or passages their users need.
Kimbodo Engineering Perspective
We would treat these releases as choices within a retrieval design, not reasons to replace an existing stack. LlamaIndex, LangChain, and Haystack can orchestrate ingestion and retrieval; Weaviate, Pinecone, Qdrant, Milvus, Elasticsearch, and Vespa can serve different indexing and search requirements. The selection should follow measured relevance, filter behavior, latency, operational burden, and access-control needs.
Fast-search map fields are most compelling when key-value filters are frequent and selective. They introduce an index to maintain and memory use to monitor [1]. Likewise, a multimodal embedding model is useful only if evaluation shows that it improves retrieval enough to justify its ingestion and query costs.
How We Would Implement It
- Define retrieval tests first: collect representative questions, expected documents or passages, permission boundaries, and latency targets. Measure both retrieval quality and answer quality.
- Build a traceable ingestion pipeline: extract text and images where needed, retain document and page identifiers, record versions, and re-index changed content. Keep metadata needed for authorization and filtering alongside each searchable unit.
- Use complementary retrieval paths: compare lexical, vector, and hybrid search; apply structured filters where appropriate; then rerank a limited candidate set before generation. Require citations back to the retrieved material.
- Test specialized indexes: for variable product attributes in Vespa, benchmark fast-search-map-field against the existing map query on production-like data, including index memory and update behavior [1]. For multilingual or image-bearing documents, compare Embed 5 Pro and Fast against the current embedding baseline using the same evaluation set [2].
- Roll out incrementally: shadow queries or route a small share of traffic to the new retrieval path, inspect failures, and retain a rollback path.
Risks, Costs and Security
Additional indexes and embeddings increase storage, ingestion work, and re-indexing costs. Faster candidate lookup does not guarantee better answers: poor extraction, chunking, permissions, or reranking can still surface the wrong evidence. Vespa exposes memory monitoring for fast-search map fields, which should be part of capacity testing [1].
Enforce document-level access controls before content reaches the generator, and test for cross-tenant leakage. Treat retrieved text as untrusted input to limit prompt-injection risk. For documents sent to an embedding service, review data residency, retention, and contractual requirements before production use.
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our RAG Development Services practice, or Estimate My RAG System.
