Documents & indexing

Supported inputs#

The dashboard accepts plain text, Markdown, and PDFs through the corresponding document operations. Text and Markdown use JSON admission; PDFs use the file endpoint.

PDF upload creates a parsing job. Parsed text then follows the existing indexing workflow. OCR quality depends on the original document; inspect the extracted source when a result looks unexpected.

Versions and publication#

A document ID identifies the document. Each accepted update creates an immutable version and an indexing receipt.

A worker prepares chunks and embeddings, then publishes the version after the required checks succeed. Until publication, an upload receipt alone does not make new content available to search.

Upload → Parse PDF, when needed → Chunk → Embed → Index → Publish

Follow the work#

Parse jobs and indexing jobs have distinct identities. Keep the returned receipt and inspect the matching job.

Recovery is allowed only when the server classifies it as safe. An uncertain provider outcome can require operator review instead of automatic replay.

Create a text document#

Send this illustrative JSON body to POST /api/documents on the Go API. Supply a fresh UUID in the Idempotency-Key header for a new operation.

{
  "id": "support-guide",
  "contentType": "text/markdown",
  "content": "# Support hours\nMonday to Friday, 09:00–17:00 UTC."
}

Deleting a document#

Deletion creates a tombstone. Deleted content is excluded from current retrieval, and saved evidence access is checked again.

Deletion does not promise physical erasure of retained historical bytes. A saved answer can become unavailable when access to its original evidence is revoked.