2.5 Vector Index Scale, Migration & Retrieval Evaluation

2.5 Vector Index Scale, Migration & Retrieval Evaluation

Scale and index operations

Approximate nearest-neighbor methods trade some recall for speed. HNSW graph parameters affect memory, build time, and search recall. Inverted-file approaches depend on training and probe choices. Managed services hide parts of this machinery, but operators still need to monitor latency, throttling, storage, index health, ingestion lag, and recall on canary queries.

Sharding and multi-index designs should follow access patterns. Separate indices can isolate departments or languages, but excessive fragmentation complicates routing and wastes baseline capacity. A shared index reduces duplication but requires robust metadata filtering and careful tenant testing. Benchmark both the average and worst-case filter selectivity because restrictive filters can change search behavior.

A repeatable evaluation workflow

Create a labeled set of queries with relevant document IDs and deliberately confusing near-matches. Compare embedding models, dimensions, chunking choices, distance functions, and index parameters using recall at k, mean reciprocal rank, normalized discounted cumulative gain, query latency, ingestion cost, and index size. Use exact search on a representative subset as a reference where practical.

Run evaluation again when documents, languages, embedding versions, or preprocessing rules change. Do not mix vectors from incompatible embedding models in one field. A model migration normally requires a new index or side-by-side vector field, backfill, validation, traffic cutover, and rollback plan.

The exam's best answer is rarely 'choose the largest vector.' Choose the configuration that meets measured retrieval quality, scale, latency, language, modality, security, and cost requirements, and preserve the evidence needed to repeat the decision.

Operational incident example

Assume query latency rises only for a tenant whose filter matches a very small fraction of a shared index. The foundation model is not the first suspect. Compare unfiltered and filtered vector latency, candidate counts, shard distribution, and index parameters. The remedy may be a tenant-specific index, a routing partition, different filterable metadata, or tuned search breadth. Increasing generation capacity will not repair vector search.

Monitor ingestion lag, failed writes, index size, search latency, recall canaries, hot partitions, and rejected queries. For graph indices, recall can drift after large update patterns or parameter changes even when the service is healthy. Schedule evaluation, not merely availability checks.

An embedding migration should be reversible. Create a new field or index, backfill from canonical text, run the labeled suite, shadow queries, and compare result sets. Cut over a small traffic percentage and retain the old index until quality and operations stabilize. Prevent new documents from entering only one index during the transition by dual writing or replaying a versioned change log.

For multi-index routing, define a deterministic directory from tenant, language, domain, or security class to eligible indices. Query fan-out needs a cap and a merge method. Returning the top score from incomparable embedding spaces is not valid unless scores have been calibrated.

Metadata and scale tradeoff

High-cardinality filters, broad wildcard fields, and large metadata payloads can raise cost or reduce performance. Keep searchable metadata purposeful and move verbose display data to a source store referenced by ID. Test mixed workloads: bulk ingestion can contend with search resources even when average query latency looks healthy.

For freshness, measure the time from source commit to searchable vector and the time from deletion to non-retrievability. For reliability, sample known query-document pairs continuously. Availability without recall can leave a vector service technically green while the application is functionally broken.

Capacity limits are safety controls as well as cost controls. A low maximum can reject growth; an unlimited or overly high setting can create an unexpected bill. Set limits from load evidence, alarm before them, and document what the application does when search capacity is exhausted.

Disaster recovery

Keep canonical text and metadata outside the vector index so the index can be rebuilt. Document encryption keys, network policies, schemas, embedding version, and replay order. Test restoration with retrieval canaries; a restored collection that answers health checks but lacks current embeddings is not recovered.

Cost attribution

Attribute embedding calls, vector storage, indexing capacity, query capacity, and reingestion work by environment and corpus. A dimensionality change can lower storage while increasing migration cost; an index split can improve isolation while duplicating baseline capacity. Compare total lifecycle cost with measured quality.

Delete abandoned test indices and revoke their access policies after experiments. Retaining stale corpora creates cost and a data-exposure path, especially when their metadata predates current authorization rules.

Vector migration gate