Importer Limits and Quotas
The concurrency, size, timeout, and provider limits the Importer service enforces, and which of them you can change
This page lists the limits the Importer enforces. They fall into three kinds:
- Hard: the request is rejected or the job fails. You cannot raise them.
- Soft: a default you can override, or that your operator can change at install time.
- Fixed: a built-in value that is not exposed as a setting.
Concurrency and jobs
| Limit | Value | Kind |
|---|---|---|
| Concurrent imports or reclusters per replica | 1 | Hard |
| Retained completed jobs per replica | 100 | Fixed |
Concurrent chat calls (CHAT_MAX_CONCURRENT_CALLS) | 16 | Soft |
Concurrent embedding calls (EMBEDDING_MAX_CONCURRENT_CALLS) | 16 | Soft |
A single lock serializes both import endpoints and POST /v1/recluster on a
given replica, whichever partitions they touch. Completed jobs are held in
memory, so they are lost when the pod restarts. See
Asynchronous import lifecycle.
Request size
There is no application-level rate limiting, and no limit on the number
of files or the total size of a POST /v1/import-multiple request.
In practice, a large inline upload is bounded by the memory available to the service, which defaults to 256 MiB at the HTTP layer and 2 GiB for the import itself, and by the 15-minute request timeout of the platform route. Neither is an install parameter.
file_id instead of embedding them as base64
file_content. The Importer then resolves them from the File Manager without
the bytes passing through the request at all.Providers
customrequires a non-empty API URL for that side (chat_api_urlorembedding_api_url).customchat additionally requires an explicitchat_model.- OpenAI-compatible providers require at least one API key.
- Image descriptions require an OpenAI-compatible vision endpoint. See Vision model requirements.
Graphs and sharding
| Limit | Value | Kind |
|---|---|---|
shard_count for a new SmartGraph or sharded Enterprise Graph | exactly 1 | Hard |
smart_graph_attribute | must be partition_id, which must also be set | Hard |
is_disjoint on a new Enterprise Graph | must be false | Hard |
partition_id length | 1-254 UTF-8 bytes | Hard |
partition_id characters | No whitespace, no :. Allowed: A-Z a-z 0-9 _ - . @ ( ) + , = ; $ ! * ' % | Hard |
| Replication factor | 3, or the cluster’s minReplicationFactor if that is higher | Fixed |
Multi-shard SmartGraphs are not supported in this version, because ArangoDB
trains the vector index per shard. For a graph that already exists,
shard_count is ignored. See
SmartGraph and sharding.
Chunking
| Parameter | Default | Rule |
|---|---|---|
chunk_token_size | 1024 | Must not be smaller than chunk_overlap_token_size. Equal values are accepted |
chunk_overlap_token_size | 128 | |
chunk_min_token_size | 64 | Smaller chunks are merged with their neighbors |
batch_size | 1000 | ArangoDB insert batch size |
entity_extract_max_gleaning | 1 |
Vector index
embedding_dimdefaults to512and has to match the embedding model you configured. See LLM Configuration.vector_index_metricacceptscosine(default),innerProduct, andl2. Any other value is rejected.vector_index_n_listsmust be a positive integer when set. When omitted, it is computed as8 * sqrt(collection_size)and capped at the collection size. It is ignored whenvector_index_use_hnswistrue.
Document conversion
| Limit | Value | Kind |
|---|---|---|
| Markdown accepted per document | 64 MiB | Soft. A larger result excludes that file as MARKDOWN_TOO_LARGE |
Which formats can be converted is not an Importer limit. See Format support.
Images
| Limit | Value | Notes |
|---|---|---|
| Local image size | 20 MiB | Larger images are skipped, not failed |
http:// and https:// image fetches | Allowlist-gated | Only hosts on the operator’s allowlist (IMAGE_FETCH_ALLOWED_HOSTS) are fetched. If no allowlist is set, only File Manager URLs are fetched |
File Manager and s3:// artifact routes | Not allowlist-gated | Read through the platform’s storage path instead of being fetched by URL, so the host allowlist does not apply |
data: URIs | Stored as-is | Inline base64 images are kept unchanged |
| Any other URL scheme | Refused |
Timeouts and retries
| Setting | Default | Kind |
|---|---|---|
ArangoDB request timeout (ARANGODB_REQUEST_TIMEOUT_SECONDS) | 900 s | Soft |
| Platform route timeout | 15 min | Fixed |
| Chat request timeout | 600 s, with a 120 s connect timeout | Fixed |
| Embedding request timeout | 120 s | Fixed |
| Vector index training wait | up to 3600 s | Fixed |
Provider retry attempts (OPENAI_MAX_RETRY_ATTEMPTS) | 10, where 0 means unlimited | Soft |
Provider retry window (OPENAI_MAX_RETRY_DURATION_SECONDS) | 3600 s | Soft |
Image upload retries (IMAGE_UPLOAD_MAX_RETRIES) | 3 | Soft |
Token limits
CHAT_MAX_COMPLETION_TOKENSis derived from the chat model’s context window, falling back to8192.EMBEDDING_MAX_INPUT_TOKENSdefaults to8192.
See Token budget for chat models for the full set of chat controls.
Project naming
The project name is not validated by the Importer, but it is used verbatim as the prefix of every collection it creates. It therefore has to satisfy ArangoDB’s collection naming rules:
- It must start with a letter or an underscore (
_). - It may only contain letters, digits, underscores (
_), and hyphens (-). - It must not exceed 256 characters, including suffixes such as
_Documents.
An invalid name is not caught at startup; collection creation fails at runtime. See Prerequisites.
Related references
- Parameters: Complete request parameter reference.
- Error Handling: Failure markers, per-file exclusion reasons, and troubleshooting.
- Reference index: Endpoints and recommended call sequence.