Upload a policy document

The document is screened, split into chunks, embedded, and written to the vector index — the same pipeline the seeded policies went through, with the same counters reported at the end. Nothing is written to disk: an uploaded document lives in memory for this process, and POST /reset clears it.

Synthetic or internal test documents only. Every upload is screened before it is read. A policy document may carry a department phone number or a mailbox in its footer — that is furniture, and it is allowed here even though the claim upload refuses it. Anything that identifies a person — a Social Security number, a date of birth, a medical record number — is refused on both.

Upload a file

.txt and .md, up to 256 KB. A form feed (\f) between pages becomes a page boundary, which is what a citation on a finding points at.
The platform reads PDF, DOCX, XLSX, PPTX and HTML and OCRs scans through Azure Document Intelligence, Google Cloud Vision or Textract. This demo has no PDF parser and no network, so it reads text.

Or paste the policy

A CPP-nnn and a vN.N in the filename or the first page attach the chunks to that policy in the library. Without them the document is still indexed and searchable, just not attached.

This document is synthetic or internal test material. It contains no protected health information and no real member record.

Cancel

What will run

Stage names are the platform's own, from its document_ingestion workflow.

storage_list_files

List the files in the dataset's source path

document_reader_unified

Read each file — PDF, DOCX, XLSX, PPTX, HTML on the platform; .txt and .md here

metadata_extractor

Pull filename, page count, digest and the policy the document belongs to

text_splitter_v2

Split each page into 1000-character chunks overlapping by 200

batch_embedder

Embed every chunk — local-hashing-256 at 256d here, text-embedding-3-small at 1536d on the platform

zip_merge

Join each embedding back to its chunk and its metadata

batch_vector_upsert

Write the points, each stamped with organization, project, dataset, document, version and is_active