Where your uploads live

For underwriting work where the documents are sensitive (rent rolls, P&Ls, OMs under NDA), this page documents the infrastructure path your uploads travel.

At upload

When you drag a document into a property's Documents section:

1. Your browser POSTs the file to relm.ai over TLS 1.2+.
2. The file is streamed directly to encrypted object storage in our US region. It's encrypted at rest with AES-256-GCM.
3. A reference (URL + metadata) is written to our application database.

The file does not transit a third-party-hosted intermediate during upload — it goes from your browser to our object storage via our edge.

During indexing

After the upload completes, the indexing pipeline:

1. Reads the file from object storage.
2. Extracts text — for PDFs, this may invoke an OCR service for scanned content. The bytes for OCR may be sent to a cloud OCR provider, covered under enterprise APIs with no training-data retention.
3. Chunks the text into searchable passages.
4. Embeds each chunk via a production embedding API (no training, contractual).
5. Writes the embeddings into our vector index, scoped to your organization.

The original file stays in object storage.

During chat / agent work

When a chat query or agent run needs a document's content:

1. The agent retrieves relevant chunks from the vector index.
2. Chunks are passed to a frontier model provider along with your query.
3. The model's response is streamed back to your browser.

The provider sees only the chunks needed to answer the query, not your entire document corpus.

At deletion

When you delete a document:

1. The object-storage file is deleted.
2. The metadata in the application DB is removed.
3. The vector-index chunks are removed.
4. Operational backups (which include the file) age out within 30 days per our backup rotation.

After 30 days, no copy of the document remains in our systems.

Cross-organization isolation

• Object-storage paths are organization-scoped.
• Vector-index namespaces are per-organization.
• Application DB queries are scoped via the org-id filter at the API layer.

There's no path for one organization's chat / agent run to retrieve another organization's document chunks.

What we don't do

• We don't ship your documents to any third-party indexer or search provider beyond our own vector index.
• We don't allow contractor / personal use of your data.
• We don't include your data in any shared dataset or benchmark.

What's next

• Privacy & data handling
• AI in Relm — what we use, what we don't
• SOC 2 posture