Some extraction jobs run in the background. Others happen while a user is waiting, or inside an agent workflow that cannot take the next step until the result comes back.
In those cases, extraction speed is not just a nice-to-have. You need high-quality structured data, but you need it fast enough that extraction does not become the bottleneck for the rest of your application.
Today we are releasing Turbo in beta, our fastest extraction tier yet, built for those workflows. On ExtractBench, it ran roughly four times faster than Cost Effective, with a median processing time of 3.7 seconds per page and a Value F1 of 0.84 (the benchmark’s measure of value accuracy and completeness). It is live now in Extract.
The fastest system we evaluated on ExtractBench
Turbo removes the separate parsing pass used by our other tiers. It extracts directly from the document pages and processes them in parallel.
We ran Turbo, our other tiers, and the fastest documented modes from several major providers on ExtractBench, our open benchmark for schema-guided document extraction. Turbo had the lowest median processing time of any system we evaluated.
Extend, the closest system in accuracy, took three and a half times as long per page.
Length is where the latency difference shows the most. Short documents run at about 3.7 seconds per page. On medium and long documents, Turbo dropped to about half a second per page because the fixed cost of a job spreads across pages processed together. Past about 16 pages, every other system's latency climbs steeply. Turbo's barely moves.
Where it sits
Use Turbo when extraction sits directly in the response path and every second matters. For background jobs where unit cost matters more than response time, Cost Effective remains the better fit. For the hardest documents and schemas, use Agentic or Agentic Plus.
What people are building
- Auto-filling forms from uploads. A user uploads an invoice and the fields fill in while they watch. Turbo returns the extraction fast enough to keep the review-and-confirm flow in one sitting.
- Keying in purchase orders. A distributor receives purchase orders as email attachments, and someone retypes every part number and quantity into the order system. Turbo pulls the line items out in seconds, while the reply to the customer is still being written.
- Letting agents act on documents. An agent that has to read a contract or a claim form before its next tool call is only as fast as the extraction. Turbo returns the structured data in seconds, so extraction does not hold up the rest of the task.
What to keep in mind
Turbo currently supports fewer input types and configuration options than the parse-backed tiers. The docs have the current list.
Try it
bash
curl -X POST 'https://api.cloud.llamaindex.ai/api/v2/extract?project_id={PROJECT_ID}' \
-H 'Content-Type: application/json' \
-H "Authorization: Bearer $LLAMA_CLOUD_API_KEY" \
-d '{
"file_input": "{FILE_ID}",
"configuration": {
"tier": "turbo",
"data_schema": {
"type": "object",
"properties": {
"vendor": {"type": "string"},
"invoice_total": {"type": "number"}
}
}
}
}'The response comes back with the schema filled in:
json
{
"data": {
"vendor": "Acme Industrial Supply",
"invoice_total": 4820.00
}
}Turbo is available today for 35 credits per page, all-in. Try it on a latency-sensitive workflow and send us your feedback and the issues you find.
