Sebastian Schkudlara
Uploading a PDF to an AI tool feels like the work is finished.
Often it is where the trouble starts.
The document may be a scan. The table may be stored as positioned text. A slide deck may look obvious to a person and arrive as a pile of disconnected fragments. Audio and video need transcription before a text model can do anything useful with them.
The file is present, but the agent still does not have reliable context.
Markdown is a useful handoff format
File2MD converts local files or remote file URLs into a Markdown-oriented set of outputs.
The current interface handles PDFs, common image formats, CSV and JSON, text files, and media such as audio, video, and YouTube links. The processing path changes with the source. A PDF can use page-level OCR. An image goes through visual text extraction. Audio and video are transcribed. Plain text can pass through with optional cleanup.
The result is not trapped inside a chat session. An agent can search it, diff it, cite it, split it into chunks, or store it beside the project that uses it.
That inspectability matters to me more than another clever upload widget.
One endpoint, two common workflows
For a local file, an app sends multipart form data to /api/file2md. For a file already online, it sends the URL instead.
The response can contain page-level Markdown, a full-document version, and processing metadata. The exact files depend on the input type.
An agentic workflow can keep the sequence simple:
- Receive a file or URL.
- Convert it to Markdown.
- Show the extracted text for review.
- Send the approved output to search, RAG, or another agent.
The conversion step stays separate from the reasoning step. If the extraction is wrong, you can inspect the bad Markdown instead of wondering why the final answer feels strange.
BYOK is the honest boundary
File conversion uses OCR, vision, or transcription services, so the hosted workflow requires a SwitchAI API key. The web interface stores that key in the browser and sends it with the request.
For an application, the key belongs in a secret store or environment variable, never in source code. Sensitive documents also need an explicit data-handling decision before they leave the machine.
That should not be hidden behind a cheerful upload button.
For private material, run a local extraction path or use infrastructure whose retention and processing rules you have accepted. For ordinary product manuals, public reports, and user-provided documents, the hosted endpoint removes a lot of integration work.
OCR still needs review
No converter can infer every table, handwriting sample, or damaged scan perfectly.
The useful product decision is that the output remains editable. A person or a second agent can correct a heading, remove a repeated footer, or reject a page before it reaches long-term memory.
That is the pattern I want in production AI: extraction first, inspection second, reasoning third.
File2MD does not make a document true. It makes the document available in a form the rest of the workflow can actually examine.