Skip to main content
File ingestion uses a two-step presigned upload so document bytes move directly to MinIO while the gateway retains tenant-scoped metadata and authorization.
1

Request an upload

In Knowledge → Static, choose a file and provide its title/catalog information. The gateway validates the request and returns a presigned upload target plus a file key.
2

Upload directly to storage

The console sends the file to MinIO using the presigned target. The API process does not proxy the whole document body.
3

Confirm the document

The console confirms the upload. The gateway verifies metadata and enqueues the document for ingestion.
4

Monitor status

Status progresses through queued and indexing to indexed, or to failed with an error message. Use Reindex after correcting a transient failure.

​
Extraction behavior

The agent streams objects from MinIO and uses source-specific parsers: pdf-parse for PDF and Mammoth for DOCX. Text is normalized and segmented before semantic chunking, then embeddings are generated in bounded concurrent batches.
Scanned image-only PDFs need OCR before upload; the current parser extracts embedded text and does not provide an OCR pipeline.

​
Troubleshooting

Verify that gateway and agent use the same bucket and credentials, the object key was not changed, and the public/presigned MinIO URL is reachable from the browser.
Check whether the PDF contains selectable text, whether the DOCX has content in unsupported embedded objects, and inspect reported word/chunk counts.
Confirm provider credentials, model availability, vector dimensions, and that Qdrant is healthy. Retry only after resolving the underlying mismatch.
Deleting a source removes its managed metadata and should remove its associated vectors/object according to the gateway service. Confirm backup and retention expectations before deletion.
Last modified on July 17, 2026