Limits

Use this page to size requests against the limits enforced by the LambdaDB service, including native embeddings, reranking, and facets. Plan quotas apply separately from per-request limits. KiB, MiB, and GiB mean 1,024, 1,048,576, and 1,073,741,824 bytes. Text byte limits use UTF-8. Character limits for document IDs, descriptions, metadata tag values, and keyword indexing count UTF-16 code units, so some characters, such as emoji, count as two.

Requests and documents

Resource	Limit	Behavior and guidance
API request body, including regular upsert and update	6,029,312 bytes (5.75 MiB), including the JSON envelope	Larger bodies return HTTP `413`. Split batches or use bulk upsert
Individual document	5 MiB of serialized JSON	Oversized documents are rejected. An update must also fit after merging with the stored document
Document ID	512 characters	Longer IDs are rejected. Assign stable IDs when you need to fetch, update, delete, or retry documents
Bulk upload object	200 MiB by default	Enforce the returned `sizeLimitBytes`; an oversized object is rejected when completing the import
Gateway request processing	29 seconds	Expiry returns HTTP `504`. A timed-out write may already have been applied; verify its outcome before retrying

The bulk limit applies to the uploaded object; individual documents must still fit their own limit. Bulk processing is asynchronous and continues after acceptance. Bulk upsert never generates native embeddings: a document with a present embedding source must include its vector.

Collections and schema

Resource	Limit	Behavior and guidance
Collections	No collection count limit	Usage quotas and request throttling still apply
Collection name	3–52 characters: ASCII letters, digits, `_`, `-`	Invalid names are rejected
Collection description	255 characters	Longer descriptions are rejected
Collection metadata tags	5 entries; keys 1–63 characters; values 1–127 characters	Keys use ASCII letters, digits, `_`, `.`, `-`; values must be nonblank and cannot contain `:`, `#`, or `,`. See metadata tags
Explicit `partitionConfig.numPartitions`	2–1,024	Applies when creating a partitioned collection; omit `partitionConfig` for an unpartitioned collection
Snapshot retention	1–31 days; default 30	No unlimited-retention setting. Existing data-versioning tags keep their pinned snapshot beyond this window
Dense vector dimensions	64–4,096	Must match the configured field. Native embedding models can impose narrower choices; see index types and supported models
Indexed keyword value	4,096 characters per value	Longer values remain stored but are omitted from the keyword index, including sorting and facets. This is not a general text-field length limit

Schema updates are additive: preserve every existing field and its configuration. See collection update rules.

Search and reads

Resource	Limit	Behavior and guidance
Query results, `size`	1–1,000; default 10	`size: 0` is allowed only with nonempty facets
Vector candidates, `knn.k`	1–1,000	Independent of the final result count and rerank candidate budget
Fetch IDs	1–1,000 per request	Split larger ID sets into fetch requests
Document list page	1–1,000; default 100	Follow `nextPageToken` to list more documents
Collection list page	1–100; default 10	A page size is not a collection count limit
Boolean query clauses	4,096	Queries exceeding the clause limit are rejected
Sparse query vector	1–4,096 term-weight entries; weights 0–64	Applies to query input; see sparse vector search
Bayesian candidate budget	`1 <= size <= candidateSize <= 1000`	Exactly two unboosted signals. With reranking, use `rerank.candidateSize` instead; see Bayesian request construction
Facet request	Up to 5 fields; 1–100 returned buckets per field, default 10	Aggregation also has distinct-value and memory budgets; see below
Facet aggregation	10,000 distinct `(field, value)` buckets and 256 KiB of distinct value bytes across requested fields	The same value on two fields counts twice. Exceeding either budget returns HTTP `400`; lowering returned bucket counts alone does not reduce aggregation work. See facet handling
Consistent-read pending payload	Less than 1 GiB by default, on the selected branch	Reaching the configured threshold returns HTTP `429`; wait for indexing to catch up

Consistent reads are available only for query and fetch against a branch directly. They exclude pending bulk imports; tags, aliases, and list requests do not support them. Large query, fetch, and list results can return isDocsInline: false and a presigned docsUrl. The inline threshold leaves room below 6 MiB for transport overhead; it is not a total-result cap. Other processing and memory budgets still apply. Follow the large-result download instructions.

Native inference

Embeddings

LambdaDB validates the following source and image limits before calling the embedding provider.

Resource	Limit	Behavior and guidance
Embedding source mappings	1–16 per vector field	OpenAI accepts exactly one text source; Cohere accepts text and image sources
Cohere input parts	16 text/image parts per input after expanding image arrays	Input is rejected when expansion exceeds the limit
Cohere images	8 per input; 1 MiB total decoded image bytes per input; 4,194,304 pixels per image	Applies to document and query inputs; see image input rules

Provider token and model context limits also apply. See text input limits for those budgets and how to split long inputs.

Reranking

Resource	Limit	Behavior and guidance
Candidate pool	`size <= rerank.candidateSize <= 1000`, with positive `size`	Default candidate budget is `max(50, size)`; each vector leg’s `knn.k` remains independent
Selected text fields	1–8 unique stored scalar text paths	Select only the fields needed for reranking
Query text	8 KiB UTF-8	Oversized input is rejected, not truncated
Rendered candidate text	16 KiB UTF-8 per candidate	Reduce selected fields or stored text
Combined query, candidate text, and custom criteria	256 KiB UTF-8	Reduce candidate count or text size
Custom criteria (`typesafe`)	2–10 distinct nonblank descriptions; 2 KiB each and 8 KiB total	Provider model context limits also apply

See native reranking for supported inputs, provider restrictions, and failure handling. Input validation and inference quota failures are not bypassed by onFailure: "returnOriginal".

Quotas and throttling

Monthly usage allowances are separate from per-request limits. See Understanding costs for Free plan read, write, storage, and inference quotas. Moving to a paid plan does not remove document, request, or model input limits. HTTP 429 can mean request throttling, too much pending data for a consistent read, temporary facet memory exhaustion, or provider-call admission exhaustion. Follow the error message: waiting for indexing differs from reducing concurrency. Honor Retry-After when present. An individual facet request over its memory budget returns 400; an inference quota failure returns 403. See Errors and retries. For tool-specific caps, see CLI limits and the MCP transport limit. Client limits apply in addition to service limits.

Get started

Migrations

Collections

Data versioning

Documents

Search

Manage Costs

Security

Requests and documents

Collections and schema

Search and reads

Native inference

Embeddings

Reranking

Quotas and throttling

​Requests and documents

​Collections and schema

​Search and reads

​Native inference

​Embeddings

​Reranking

​Quotas and throttling

Requests and documents

Collections and schema

Search and reads

Native inference

Embeddings

Reranking

Quotas and throttling