Document Storage

Tenant Isolation

Each tenant has its own storage bucket, named by the tenant’s unique ID. This per-tenant isolation ensures:

Clean data separation between tenants
Simple bulk deletion when a tenant is removed
Independent access control at the bucket level

Storage Layout

All storage paths use a flat structure based on the document and version IDs, deliberately independent of your folder hierarchy. This means moving a document between folders never requires changing storage paths.

{tenant_id}/
  documents/{document_id}/{document_version_id}/
    source.{pdf,docx,pptx,md,txt}       # Original uploaded file
    cleaned_source.pdf                    # Watermark-removed PDF (if applicable)
    standard_pipeline.json                # Structured conversion output
    page_screenshots/                     # Full-page screenshots (WEBP)
      p1.webp                            #   Page 1
      p2.webp                            #   Page 2
      ...
    images/                              # Extracted image crops (WEBP)
      0.webp
      1.webp
      ...
    tables/                              # Extracted table screenshots (WEBP)
      0.webp
      1.webp
      ...

What Gets Stored

Source files

Your original uploaded document is stored as-is under the source.* key, preserving the original file extension.

Cleaned PDFs

If your PDF contains watermarks, the preparation step produces a cleaned version at cleaned_source.pdf. The original source is preserved. See PDF Watermark Removal for details.

Conversion output

The document conversion step produces a structured JSON representation of your document’s content, stored as standard_pipeline.json. This intermediate format is used by the chunking step.

Visual assets

All visual assets are stored as WEBP images (quality 85, DPI 144):

Asset Type	Path Pattern	Description
Page screenshots	`page_screenshots/p{N}.webp`	Full-page renders, 1-indexed
Images	`images/{N}.webp`	Extracted image crops from the document
Tables	`tables/{N}.webp`	Screenshots of detected tables

Page screenshots capture each page of the document as a high-quality image. Image and table crops are extracted from the page images based on detected bounding boxes during conversion.

Storage Operations

The platform supports these storage operations:

Operation	Description
Upload	Store raw bytes at a given path
Download	Retrieve object content
List	List objects by path prefix
Delete	Batch delete objects
Presigned URLs	Generate time-limited download links

URI Format

All internal storage references use the s3://{bucket}/{key} URI format, making it easy to locate any stored asset.

Get Started

SDKs & MCP

Cookbook

Concepts

Ingestion Pipeline

Agent

Infrastructure

Design

Operations

Tenant Isolation

Storage Layout

What Gets Stored

Source files

Cleaned PDFs

Conversion output

Visual assets

Storage Operations

URI Format

​Tenant Isolation

​Storage Layout

​What Gets Stored

​Source files

​Cleaned PDFs

​Conversion output

​Visual assets

​Storage Operations

​URI Format

Tenant Isolation

Storage Layout

What Gets Stored

Source files

Cleaned PDFs

Conversion output

Visual assets

Storage Operations

URI Format