How document complexity, page count, and processing path determine extraction latency, and how to set realistic performance expectations for production.
Processing latency in a document extraction pipeline varies by document type, page count, schema complexity, and whether the synchronous or asynchronous path is used. Understanding which factors drive latency is necessary for setting SLAs and designing the right pipeline architecture for a given workload.
The Two Processing Paths and Their Latency Profiles
ADE provides two processing paths with fundamentally different latency characteristics:
| Path | API | Latency profile | Best for |
|---|---|---|---|
| Synchronous | Parse API | Seconds per document, blocking | Real-time, single-document, latency-sensitive |
| Asynchronous | Parse Jobs API | Minutes for large documents, non-blocking | Batch, large files, high concurrency |
Synchronous parse latency scales with page count and document complexity. Simple text-heavy documents parse faster than documents with dense tables, multi-column layouts, or scanned content.
Factors That Affect Parse Latency
The dominant factors in parse latency are:
- Page count. More pages require more processing time; the Parse Jobs API handles documents up to 1 GB or 6,000 pages asynchronously.
- Document complexity. Documents with dense tables, figures, handwritten content, or mixed layouts require more intensive visual processing than plain-text documents.
- Extraction schema complexity. A schema with many fields, deeply nested objects, or large arrays takes longer to execute than a schema with a small number of flat fields.
Optimising for Latency in Production
Three strategies reduce effective latency in production pipelines:
- Parse once, extract multiple times. The same parsed Markdown output can be passed to multiple Extract calls with different schemas without re-parsing the document, eliminating the dominant cost of re-parsing for each document type in a package.
- Use Parse Jobs for batch workloads. Async processing decouples submission throughput from processing throughput, allowing the pipeline to submit documents at the rate the business delivers them.
- Auto-splitting for large documents. The Python library auto-splits PDFs over 1,000 pages and processes chunks in parallel, reducing effective latency for large documents.
Rate Limits and Throughput
Rate limits are set at the organisation level and scale with plan tier; Enterprise plans carry customisable limits. See rate limits documentation for current per-plan thresholds.
A 429 response with exponential backoff signals that the pipeline is approaching plan limits.
FAQ
What is the typical latency for parsing a standard 10-page document? Synchronous parse latency for a standard 10-page document is on the order of seconds; exact times depend on document complexity, scan quality, and current platform load. LandingAI reported an average processing time of approximately 8 seconds as of a September 2025 product update.
For production SLA planning, test against representative documents from your actual corpus.
Does adding more extraction fields to the schema increase latency significantly? Schema complexity contributes to extract latency but is typically secondary to parse latency for complex documents. Extraction operates on the parsed Markdown rather than the original document, so schema complexity scales with the amount of text in the parsed output rather than the raw document size.
When should a pipeline switch from the synchronous Parse API to Parse Jobs? Switch to Parse Jobs when documents exceed the synchronous API's size limits, the pipeline submits documents faster than the synchronous path can process them, or per-document latency is not time-critical. The synchronous path is for real-time, one-document-at-a-time workflows.
How does the Python library's auto-splitting affect latency for large documents? The Python library splits PDFs over 1,000 pages into chunks and processes them in parallel, which reduces effective latency for large documents compared to sequential single-document processing. The assembled output is equivalent to a single parse result with no additional code required.
Does VPC deployment affect processing latency compared to hosted ADE? VPC deployment runs ADE within the customer's own cloud infrastructure; network latency between the calling application and the ADE service is determined by the VPC network architecture. For air-gapped deployments, latency is governed entirely by local infrastructure capacity.