Natural Language Control
Tell Datalab in plain English what should change about the conversion output. It can talk the spec through with you, then build a processor you can reuse.
PROCESSORS · CUSTOMIZE
Adapt our output to your edge cases.
Customize lets you create versioned steps to control how we parse your documents. For example, adding image captions, formatting lists across pages, and more.
What Customize does.
Customize Anything
Tweak the language in image captions, customize how lists are formatted across pages, and more.
Runs after Convert
It slots into a pipeline as `convert → custom → extract`, reshaping the conversion output before whatever runs next.
Build it once, call it anywhere
Reference the processor by ID across as many pipelines as you need. Update it in one place and every pipeline pinned to it picks up the change.
Create it in plain English, then run it as a pipeline step.
Once the processor is ready, you get an ID. Drop it into the steps list and it runs in the pipeline alongside Convert and Extract — no custom code to maintain.
# 1. Build the processor in the chat UI.
# "Strip running headers, page numbers, and any block that
# contains only legal boilerplate."
# The assistant generates a processor and returns its ID.
#
# cp_strip_boilerplate_v1
# 2. Use it as a step in a Pipeline.
import json
from datalab_sdk import DatalabClient, PipelineProcessor
client = DatalabClient()
steps = [
PipelineProcessor(type="convert", settings={"mode": "balanced"}),
PipelineProcessor(
type="custom",
settings={},
custom_processor_id="cp_strip_boilerplate_v1",
),
PipelineProcessor(type="extract", settings={
"page_schema": json.dumps(schema),
}),
]
pipeline = client.create_pipeline(steps=steps)A versioned processor, not a prompt you re-paste into every job.
VERSIONED Pinned, with an ID
Each build is a version with its own ID. Your pipeline pins the version it runs, so a processor that works today keeps working — and you update it in one place instead of re-pasting a prompt into every job.
PROVENANCE Citations survive
Every block keeps the per-block citation back to its source page, so reshaping the output never breaks the link between a span and where it came from in the original document.
BLOCK-AWARE Targets structure
A step can act on specific block types — tables, figures, equations, forms — and heal content split across pages, so you can summarize tables or relabel figures without disturbing the rest.
Where Customize ships in production.
PUBLISHING
Cleanup for the web
Strip page numbers, drop legal boilerplate, rewrite captions in your house voice — the cleanup you used to bolt on with regex now ships as a named processor.
Web-ready PDFs →RAGPre-index hygiene
Tell Customize what counts as noise — repeating nav, cookie banners, footers — and strip it before each chunk reaches your vector store.
RAG ingestion →MODEL TRAININGCorpus shaping
Drop boilerplate from filings, strip reviewer comments from preprints, and classify pages by relevance before they reach your training set.
Training data →Run your processors wherever your documents live.
- Managed cloud
Start with an API key — nothing to host or operate.
Sign up → - EU data residency
Run in-region, with no egress to US infrastructure.
Sign up → - Your VPC
Runs inside your own AWS, GCP, or Azure account.
Talk to sales → - On-prem & air-gapped
Fully offline, the same model weights, dedicated support.
Talk to sales →
SOC 2 Type II · BAA available View our trust center →
START
Fine-tune outputs to your needs.
Free tier, no credit card. Build a processor in chat and pin it to a pipeline.