What tool structures unpublished scientific data so AI can analyze it?
The four steps, and the good tools that do them
Each one is good at its job. They are not weak. They solve a different problem than the one you have when you paste a file into an AI and get a confident wrong answer.
Parsing. MinerU turns a PDF into machine-readable text and tables. It is a strong parser. But you get extracted text, not structured scientific fields, and not something an AI can query with any memory of what it means.
Cleaning. OpenRefine tames messy spreadsheets: dedupe, normalize, reconcile. I still use it for that. But a clean CSV is still just a CSV. Cleaning does not give it provenance or make it AI-queryable.
Packaging. ISA-API and RO-Crate describe an investigation and bundle research objects with metadata. They matter for FAIR data and reproducibility. But they describe and package your data. They do not turn it into something an AI reasons over in a chat.
Validation. Great Expectations tests your data against rules and catches quality problems. Great in a pipeline. But validation checks data. It does not represent it for a model.
The step none of them do: grounding
That gap is the one I built KORE to fill. KORE takes your raw material and produces a portable, provenanced bundle that any AI reads as a tool, over MCP. Every value is mapped to a scientific schema, so a Ct value is understood as a Ct value and not just a string. Every field carries a citation back to the exact passage it came from, so the model can tell you what its answer is based on. That is the grounding layer, and it is the step that sits after parsing, cleaning, packaging, and validating.
Honest caveat: KORE reads text and tables today, not images or figures yet.
Who does what
| Step | Tool | Does KORE do this? |
|---|---|---|
| Parse a PDF | MinerU | Not its job. Hand it parsed or raw, it structures the result. |
| Clean a spreadsheet | OpenRefine | Not its job. It works on messy or clean input. |
| Package with metadata | ISA-API, RO-Crate | No. Complementary, not competing. |
| Validate data quality | Great Expectations | No. Different problem. |
| Ground it so an AI can reason over it | KORE | Yes. This is the category. |
So do you still need the others?
Maybe. KORE is not a better parser or a better cleaner, and it does not replace RO-Crate. It is the step after them. Hand KORE a messy pile and it structures it. Hand it already-clean, well-packaged data and it grounds it for your AI. They prepare the data. KORE makes it usable in a chat.
FAQ
Is KORE a PDF parser like MinerU?
No. A parser gives you text. KORE gives you structured, provenanced fields an AI can query.
Is KORE a data cleaner like OpenRefine?
No. Cleaning tidies a spreadsheet. KORE represents it so a model understands what each value means.
Does KORE replace RO-Crate or ISA?
No. Those package and describe your data. KORE grounds it for AI. They work together.
Can't I just use a model with a big context window instead?
No. A bigger window still parses your file lossily, keeps no memory across sessions, and cannot tell you which row a number came from. Those are structural gaps, not size gaps.
Does KORE read images or figures?
Not yet. Text and tables today.
How do I try it?
Add KORE in Claude and ask it one question about your own data.
Built by GRAVL. Organizing the world's unpublished scientific data.
Try it on your own work
You can try this now: add KORE in Claude and ask it one question.