6. Vector Store Design and Lifecycle

Choose storage, design chunk schemas, enforce metadata filters, and manage updates, deletes, and index migrations.

Vector Store Options

Options include dedicated vector databases, search platforms with vector support, relational database extensions, and managed provider-hosted stores. Compare metadata filters, hybrid search, update/delete semantics, backup/restore, consistency, scale, regional controls, cost, and operational effort.

Managed retrieval reduces infrastructure code but may constrain customization. Self-managed options give control but add monitoring, upgrades, backups, and incident response. Keep a recoverable source of truth so an index can be rebuilt.

Implementation example

System of record: source documents, ACLs, versions, jobs
Retrieval index: vectors, chunks/references, searchable metadata
Cache: short-lived derived results with scoped keys

💡 Choose based on requirements and operational trade-offs.

Key Takeaways

Inspect the relevant pipeline stage instead of guessing.
Preserve source provenance, versions, and access scope.
Measure quality on representative queries.

Knowledge Check

What should guide vector-store selection?

Design a Chunk Schema

Each chunk needs a stable ID, source document ID, source version, text, embedding, source location, tenant/ACL metadata, and transformation versions. Stable IDs should account for document revisions and chunk boundary changes. Stale chunks must be removed or superseded on reindexing.

Metadata fields should reflect real filters and citation needs. Too few fields make authorization and provenance difficult; too many unconstrained fields add complexity.

Implementation example

{
 "chunkId":"doc-18:v3:chunk-07",
 "documentId":"doc-18",
 "documentVersion":3,
 "tenantId":"tenant-42",
 "chunkerVersion":"semantic-v2",
 "embeddingModel":"embedding-v1"
}

💡 Versioned records support updates, citations, and reproducible retrieval.

Key Takeaways

Inspect the relevant pipeline stage instead of guessing.
Preserve source provenance, versions, and access scope.
Measure quality on representative queries.

Knowledge Check

Why include document version on a chunk?

Metadata Filters and Tenant Isolation

Metadata filters constrain retrieval by tenant, group, publication status, language, date, and source type. Derive filters from authenticated server-side context; do not allow a model or untrusted query parameter to broaden access. Test how the chosen store applies filters, especially with approximate and hybrid search.

Use namespaces, partitions, row-level policies, separate indexes, or dedicated deployments according to threat model. A post-retrieval filter may be too late if unauthorized text has already reached intermediate systems.

Implementation example

function buildFilter(auth) {
  if (!auth?.tenantId) throw new Error("Missing tenant");
  return {tenantId:auth.tenantId,status:"published",deleted:false};
}

💡 The filter must be enforced before unauthorized text enters model context.

Key Takeaways

Inspect the relevant pipeline stage instead of guessing.
Preserve source provenance, versions, and access scope.
Measure quality on representative queries.

Knowledge Check

Where should tenant scope originate?

Upserts, Deletes, and Blue/Green Indexing

Index lifecycle errors create stale or unauthorized answers. A document update can leave old chunks behind; a deletion can leave vectors searchable; an embedding change can mix incompatible spaces. Use versioned chunk sets, transactional updates, or blue/green indexes.

For blue/green, build a new index, reconcile updates, run retrieval and security tests, switch traffic, retain the old index for rollback, then remove it when safe. Track status and counts throughout the process.

Implementation example

Build index B -> validate quality and ACLs -> switch traffic
            -> retain index A for rollback -> delete A when safe

💡 A staged cutover supports validation and rollback.

Key Takeaways

Inspect the relevant pipeline stage instead of guessing.
Preserve source provenance, versions, and access scope.
Measure quality on representative queries.

Knowledge Check

What is a benefit of blue/green indexing?

Vector Store Options

Design a Chunk Schema

Metadata Filters and Tenant Isolation

Upserts, Deletes, and Blue/Green Indexing

STAY UPDATED WITH BYTES!