Vector Search

Read this file when a design involves semantic similarity, RAG retrieval, recommendations, AI-agent memory, or anomaly detection over data in DynamoDB. Skip it otherwise.

Every error string quoted below is reproduced verbatim from the service, so you can match on it when diagnosing.

What belongs here, and what does not

Vector search changed where semantic similarity belongs. It changed nothing about lexical search. When a request bundles several kinds of “search,” split it:

Do not route lexical or geo search to a vector index — ANN over embeddings does not do substring, prefix or edit-distance matching. A workload needing both semantic and lexical search legitimately uses both systems.

Requirements and limits

• On-demand capacity is mandatory. Vector indexes are not supported on provisioned tables. A provisioned table must be switched to PAY_PER_REQUEST first — and note the 24-hour capacity-mode cooldown (Fact #3).
• A minimum SDK/CLI version applies. The operations ship in botocore/boto3 ≥ 1.43.64 and AWS CLI v2 ≥ 2.36.16. Below those versions the operations are absent from the client entirely: aws dynamodb search-vectors fails with Found invalid choice, and hasattr(client, "search_vectors") is False. Prefer that hasattr probe over a version comparison — it tests the thing you actually need and cannot go stale. A stale SDK is not evidence the feature does not exist — check before concluding anything, and tell the user to upgrade rather than reporting the feature as unavailable.
• Up to 5 vector indexes per table. Maximum 4,096 dimensions. TopK must be 1–100. At most one HASH search-schema element per index, and up to 18 inline filters. Responses cap at 16 MB with no pagination. These values are subject to change and quotas of this kind generally rise rather than fall, so treat them as the design envelope to stay inside, not as a fact to quote back to a user who is near a limit — confirm the current figure in AWS Service Quotas or the DynamoDB developer guide before telling anyone their design is impossible.
• One online index operation at a time per table, shared with GSI creation. Two Create actions in a single UpdateTable fail with LimitExceededException: Subscriber limit exceeded: Only 1 online index can be created or deleted simultaneously per table.
• Query, Scan, PartiQL and DAX do not work against a vector index. Query is rejected with ValidationException: Query operation not supported on this index type.
• Vectors are stored in the index at 32-bit float (f32) precision. Higher-precision values are accepted and lose precision on the way into the index.

Creating an index

Every attribute named in SearchSchema must also appear in the table’s AttributeDefinitions, exactly as a GSI key attribute must. Omitting it fails with ValidationException: One or more parameter values were invalid: One element in SearchSchema is not defined in attribute definitions.

On a new table, use CreateTable with --vector-indexes:

On an existing table, use UpdateTable with --vector-index-updates:

Delete with --vector-index-updates '[{"Delete": {"IndexName": "ProductEmbeddingIndex"}}]'. Deleting an index does not touch the base table or its items. A vector index is removed with its base table, so a table teardown needs no separate index step.

VectorAttribute is an object carrying AttributeName, not a bare string. The distance function key is DistanceFunction. There is no VectorIndexConfig, no VectorAttributeName, no DistanceMetric, and no way to configure a vector index through a GSI.

SearchSchema: the partition key is a scaling decision

A HASH search-schema element partitions the index so each SearchVectors call examines only one partition-key value’s vectors. That lowers latency and cost and scales throughput horizontally — the per-partition-key limits (1 GBps search, 10 MBps write) multiply across distinct values. Choose low-to-medium cardinality that matches how you query: a tenant id, a category, a region. Avoid a unique per-item id (each partition has no neighbours, so recall collapses) and avoid a boolean (no spread).

If you define a HASH element you must supply its value on every search. Omitting it fails with ValidationException: SearchConditionExpression must be provided when SearchSchema has a HASH key. Defining no SearchSchema at all is simpler and needs no SearchConditionExpression, but every search then examines the whole index.

INLINE_FILTER elements are optional at search time and are filtered at the storage layer.

Waiting for the index — get this right or your code is environment-dependent

The correct readiness predicate is IndexStatus == "ACTIVE" and Backfilling is not true. There is no BACKFILLING index status value.

Both obvious shortcuts are wrong, because the two creation paths report Backfilling differently:

Searching too early fails with ValidationException: Cannot search backfilling vector index: <indexName>.

Backfill duration is driven by index construction, not item count. Adding an index to a table holding six items took 8m33s in testing, against 20s for an inline CreateTable on a new table. Poll; do not assume a small table is quick.

ACTIVE is not sufficient on its own. SearchVectors is served by a dedicated search endpoint, separate from the one serving DescribeTable, and it can lag briefly after the index reports ACTIVE. During that window searches fail with ValidationException: The table does not have the specified index: <indexName> — the same message you get for a genuinely wrong index name. Treat a ValidationException on the first searches as retryable and take the first successful SearchVectors response as the real readiness signal. Code that searches immediately after observing ACTIVE works in one account and fails in another.

Writing vectors

Store the embedding as a list (L) of numbers (N) on the item, written with the ordinary PutItem / UpdateItem / BatchWriteItem / TransactWriteItems APIs. Because a vector is long, write it from a file rather than inline on the command line.

Silent de-indexing — the sharpest edge here

If a vector index defines a HASH search-schema element, an item written without that attribute — or with it later removed via UpdateItem ... REMOVE — is accepted on the base table with no error and is never replicated into the index. It will not appear in SearchVectors results even though the item and its embedding are intact and GetItem returns them normally.

This is expected behaviour, not a defect — it is the same semantics as a sparse GSI, where an item lacking the index key simply is not in the index. It bites harder here because the attribute is usually one every item is supposed to have, and the symptom is retrieval quality quietly degrading rather than a visible error.

Symptom signature. Items absent from similarity search; GetItem returns them; embeddings valid; IndexStatus ACTIVE; nothing in CloudWatch; no failed writes. When you see that combination, go straight to “which items are missing the search-schema partition key” — not to corruption, not to replication lag.

Fix. Re-write the affected items with the attribute restored, preserving the existing embedding. Re-creating an item without its vector leaves it permanently unsearchable, which is worse than the original bug. Do not drop and recreate the index: backfill re-derives from the same base-table items, so nothing changes.

Prevent it on the write path. Treat a search-schema HASH attribute as mandatory on every item that must be searchable — guard it at write time in the same spirit as the attribute_exists / attribute_not_exists guards in Patterns #2 — or define no partition key at all if you cannot guarantee the attribute is always present.

Embeddings go stale silently

DynamoDB does not recompute embeddings. Change the source content and the stored vector still reflects the old text, so the index keeps returning matches against content that no longer exists. Regenerate the embedding with the same model and write it back whenever the source changes — capturing content changes with DynamoDB Streams and re-embedding downstream is the usual shape (Integration #1 applies: make that consumer idempotent).

Searching

SearchVector is a plain JSON array of number objects — [{"N": "0.1234"}, ...] — not wrapped in a DynamoDB L. The L wrapper is only for the stored item attribute. Wrapping it is rejected client-side before any API call: Invalid type for parameter SearchVector ... valid types: <class 'list'>, <class 'tuple'>.

The query vector must come from the same embedding model and have the same dimensionality as the index. A mismatch fails with ValidationException: Input search vector dimension 3 does not match vector index dimension 4.

SearchConditionExpression supports the equality operator only. Comparison, range and set-membership operators (<>, <, <=, >, >=, IN) are rejected with ValidationException: Invalid SearchConditionExpression: Invalid comparator used in SearchConditionExpression. There is no FilterExpression or Filter parameter on SearchVectors.

TopK outside 1–100 is rejected — above the range server-side (Provided TopK value '101' is out of valid range. The value must be between 1 and 100 inclusive), below it client-side.

What comes back

Results arrive sorted most-similar-first, each with the projected attributes and a Score. Two things surprise people:

• The vector attribute is excluded by default, even under ProjectionType: ALL. Request it explicitly with ProjectionExpression if you want it — and note it is large.
• Only attributes projected into the index can be returned. Asking for one that is not projected is silently ignored: no error, the attribute is simply absent from the result. If you need more attributes, either recreate the index with a wider projection or hydrate from the base table with GetItem/BatchGetItem after searching.

Distance functions and score direction

Do not assume “higher score is a better match” — that is true only for DOT_PRODUCT, and it is the reverse for the other two. Do not assume scores are non-negative when applying a threshold. The simplest safe approach is to rely on the returned ordering rather than re-sorting by raw score.

The choice is a real ranking decision, not a formality. Measured on identical data with the identical query vector [1,0,0,0], the stored vector [10,0,0,0] ranks tied-first under COSINE (0.0) and last under EUCLIDEAN (9.0) — because COSINE ignores its magnitude and EUCLIDEAN does not. If you use DOT_PRODUCT and want direction-based similarity, normalise embeddings to unit length first; normalised, it ranks the same as COSINE.

What you cannot change after creation

Dimensions, DistanceFunction, SearchSchema (partition key and inline filters) and Projection are all fixed when the index is created. UpdateTable cannot alter any of them. Changing one means creating a new index under a new name, waiting for it to backfill, cutting reads over, then deleting the old one — the same additive-migration shape as a GSI key change (Fact #9). Because only one index operation runs at a time per table, this is sequential and slow.

Choose the embedding model before creating the index, since the model fixes the dimensionality. Validate the distance function against a representative dataset first. You can run up to 5 indexes on one table, which is the supported way to A/B two embedding models over the same data: store each model’s vectors in a different attribute, index each, and compare relevance before migrating.

Security

IAM controls vector operations as follows:

• Creating/deleting an index: dynamodb:CreateTable or dynamodb:UpdateTable on the table.
• Searching: dynamodb:SearchVectors on the index resource, ARN shape arn:aws:dynamodb:<region>:<account-id>:table/<table-name>/index/<index-name>.
• Writing items with vectors: the ordinary write permissions. Nothing extra.

Fine-grained access control does not apply to SearchVectors. The dynamodb: condition keys that enforce FGAC — dynamodb:LeadingKeys, dynamodb:Attributes, dynamodb:Select — are absent from the SearchVectors request context. A policy statement whose Condition references one of them therefore does not match, and the result is denial rather than a grant. Adding dynamodb:SearchVectors to an existing FGAC-conditioned statement silently breaks vector search while the statement’s other actions keep working normally — which makes it look like a broken index rather than a policy problem.

Grant dynamodb:SearchVectors in its own statement, scoped to the index ARN, with no dynamodb: FGAC conditions attached:

Note what that Condition is and is not. aws:PrincipalOrgID is a global condition key, so it evaluates normally and the statement still matches. The keys that must stay out of this statement are the dynamodb: FGAC ones above. Scope the Resource to the specific index rather than .../index/* — a wildcard grants search over every index on the table, including ones added later.

“No dynamodb: conditions” does not mean “no conditions”. Only the FGAC keys are missing from the request context; the standard global condition keys still evaluate normally and should be used, particularly on a multi-tenant index. aws:PrincipalOrgID and aws:PrincipalTag/* restrict who may search, aws:SourceVpce and aws:SourceIp restrict from where, and aws:RequestedRegion bounds the region. What none of them can do is restrict which partition-key values a permitted principal reaches — that is the gap, and it is narrower than “this statement cannot be conditioned at all”.

A search-schema partition key is not a security boundary. Scoping searches to one tenant’s partition-key value is a data-locality and performance optimisation. Any principal holding dynamodb:SearchVectors on the index can search any partition-key value, and because FGAC condition keys do not apply you cannot restrict which values they reach at the IAM layer. If the workload requires strict tenant isolation at the data layer, use separate tables or separate indexes with distinct IAM grants per tenant. An application layer that injects the caller’s tenant value and never accepts it from the client is a useful complement, not a substitute.

Log the searches, because IAM cannot gate them. Preventive controls have a known hole here, so the detective control carries more weight than it normally would: enable CloudTrail data-event logging on any table whose vector index serves more than one tenant. It is the only record of which partition-key value each principal searched, which is exactly the dimension IAM cannot constrain — so without it a cross-tenant search caused by a compromised principal or a bug in the application-layer injection leaves no trace at all. Pair it with a CloudWatch alarm on VectorSearchRequestBytes per principal: a caller enumerating other tenants’ partitions shows up as a volume anomaly before it shows up anywhere else. Encrypt the trail’s S3 bucket and any log group with a customer-managed KMS key when the logged partition-key values are themselves sensitive.

Rate-limit SearchVectors per principal at the application layer. Because every call returns similarity scores, a caller with legitimate search access can probe the embedding space of a partition through repeated targeted queries and infer relationships it was never shown directly. That makes the throttle a confidentiality control as well as a cost one — which matters, because a throttle filed under “cost” gets dropped from any design where cost is not a concern, and multi-tenant indexes holding valuable content are exactly where it should not be.

Encryption at rest is inherited from the base table — AWS-owned, AWS-managed or a customer managed KMS key. There is no separate vector index encryption setting.

Treat the embeddings themselves as sensitive data. An embedding is a lossy but real encoding of the content it was derived from, and similarity results leak information about neighbouring items even when no raw text is returned — an attacker with search access can learn that two records are semantically close without ever reading either. Two consequences worth acting on: the vector attribute being excluded from results by default is a security default, not just a bandwidth one, so do not add it to ProjectionExpression without a reason; and anything that logs a full item — Lambda consumers, debug traces, DLQ payloads — should be treated as handling sensitive data, with KMS encryption on the log group.

Cost

Vector index usage is metered in bytes, separately from base-table RCU/WCU, across three dimensions. Rates below are us-east-1 Standard; confirm against the DynamoDB pricing page for your region before quoting a figure.

Standard-IA is 125% on both request dimensions and 40% on storage.

Two things make search cost unmodellable from a design alone: it scales with the volume of vector data the search examines, which depends on index size and partition layout rather than on your query vector; and there is a 1 KB minimum per request for both writes and searches. That minimum is per request, not per index — one write touching several of a table’s indexes incurs it once. It also means low-dimension vectors do not meter proportionally lower: a 4-dimension vector holds 16 bytes of f32 data and still meters at 1,024 bytes.

So do not estimate search cost from dimension count. Get VectorSearchRequestBytes and VectorWriteRequestBytes from ReturnConsumedCapacity on your own workload, or from the CloudWatch metrics of the same names (dimensioned by TableName and VectorIndexName), and validate against a representative dataset before sizing.

Interaction with other DynamoDB features

• Global tables — the index definition replicates automatically to a new replica Region; you do not create it there. Replication into the index is asynchronous even under multi-region strong consistency (MRSC), so a vector written in one Region may not appear in another Region’s search results yet. Each replica’s index backfills independently. ANN means ranking can differ slightly between Regions over identical data.
• Streams — work normally and are the recommended way to trigger re-embedding.
• TTL — an expired item’s index entry is removed like any delete.
• PITR / backups — the index definition is restored and rebuilt from base-table data, so it backfills before it is searchable.
• Export / import — exports contain the vector attributes; on import, define the index in the import request as you would for CreateTable.
• DAX — does not support SearchVectors. Send searches directly to DynamoDB.

Troubleshooting