How to build a document collection in DataLens
Documents are not tables. A collection stores them whole rather than forcing them into rows — and shows you exactly what text was extracted.
What you will do
You will create a document collection, add documents to it, and preview what DataLens actually reads out of a file before committing to it.
A collection is a dataset, but it is not a tabular ingest: there are no rows to profile and no working copy to make. It gets its own source tab for that reason.
When this is useful
- You have policies, contracts, reports or notes that need to sit alongside your data.
- You are preparing source material for AI work and need to know what text was actually extracted.
- You want to check a document set for what is in it before committing to processing it.
Before you start
- Documents in a supported format — Which formats a deployment can read is a property of the deployment, not the browser. The panel shows the list your deployment supports — typically plain text, Markdown, HTML and Word (DOCX).
- Realistic expectations about PDF — PDF extraction is not enabled in every build. Where it is not, a PDF is refused rather than stored with empty content. See the limitations below.
Steps
Open Data → Connections → Documents
Select Data, then Connections, then the Documents tab.
The document upload panel on the left, and Collections in this workspace on the right.
Check which formats this deployment reads
The panel shows a format table listing what can be read, with file size and file count limits.
This list comes from the server rather than being fixed in the browser, so it is the authoritative answer for your deployment.
A table of formats with their status.
Add your documents
Select files to add. The file picker only offers extensions your deployment can actually read.
Files that cannot be accepted are refused with a reason rather than silently dropped, so you always know what did not make it.
Your documents queued, with anything refused listed separately and explained.
Preview what is actually extracted
Use the preview action on a document to extract that one file and see exactly what DataLens reads out of it. Nothing is stored by this preview.
This is the step that prevents a nasty surprise later. A document that looks fine in Word can extract very differently, and this is where you find out.
The extracted text for that document, and nothing written to the collection.
Create the collection
Name the collection and create it. The documents are stored whole — the bytes are kept, not parsed into rows.
The collection appears under Data → Datasets in the Documents section.
The new collection listed under Collections in this workspace.
What happens next
Document operations work on the collection's text — including chunking strategies for splitting documents into passages.
Where a chunking strategy depends on structure the source does not have, DataLens reports that rather than producing one enormous chunk. Read what it tells you; a warning here saves debugging later.
Example
A team loads their data policies as a collection. They preview one policy first and see the headings survived extraction, which tells them heading-based chunking will work on this set — a five-second check that would otherwise have been discovered much later.
Tips
- Always preview one representative document before loading the whole set. It costs seconds and tells you whether extraction works on these files.
- Trust the format table in the panel over any general expectation. It reflects what your deployment can actually do.
- If files are refused, read the reason. It names the problem rather than leaving you to guess.
- Documents are not rows. Do not expect profiling, cleaning or charting on a collection — that is not what it is.
Limitations
- PDF extraction is not enabled in every build. Where it is not, PDFs are refused rather than stored empty — a deliberate choice, so no document is ever recorded with no content.
- A collection is not a tabular dataset. There are no rows to profile, no working copy, and the Prepare and Analyse screens do not apply to it.
- File size and file count limits apply and are shown in the panel.
- Heading-based chunking needs headings to survive extraction. Where they do not, DataLens reports it rather than failing quietly.
Related how-to guides
Connect an AI model
Add a model provider with your own credentials, choose which models are allowed, and set a default.
Upload a CSV or Excel file
Load a spreadsheet or delimited file and get a profiled dataset you can clean, model and chart.
Generate a data document
Produce an architecture document, a model spec, a mapping or a runbook from what you have already built.
Related questions
More of these on the DataLens FAQ page.
Can DataLens read PDF documents?
It depends on the deployment. PDF extraction is not enabled in every build, because recovering text from a PDF needs a real PDF engine — a PDF stores glyph positions rather than text. Where extraction is not enabled a PDF is refused rather than stored with empty content. The format table in the Documents panel shows what your deployment supports.
What is the difference between a document collection and a dataset?
A document collection stores files whole and works on their text. A dataset is tabular — rows and columns that can be profiled, cleaned, joined and charted. Documents get their own source tab precisely because they are not rows.
Try this in DataLens
DataLens is in private beta. Request access and work through this guide on your own data.
Request beta access