docsmith / v0.2

Documents in. Markdown, text, pixels and clean images out.

A small HTTP API for the boring half of document work: turn a PDF into Markdown that a retrieval pipeline can actually chunk, pull the raw text, render pages to PNG or JPEG, and convert or shrink images. No models, no queues, no per-page surprises. One request, one answer.

Endpoints

endpoint	what it does
POST /v1/pdf/markdown	Markdown with heading levels, rejoined paragraphs, normalised lists and page boundaries. Returns the heading outline as JSON too.
POST /v1/pdf/text	Extract text, optionally layout-preserving and page-ranged.
POST /v1/pdf/info	Page count, page size, encryption, title.
POST /v1/pdf/images	Render pages to PNG or JPEG at 72-300 dpi. One page returns an image, several return a zip.
POST /v1/image/convert	JPEG, PNG, WebP or GIF, with resize, quality and metadata stripping.
GET /healthz	Liveness plus which engines are present.
GET /openapi.json	OpenAPI 3.0 description, importable into any client or marketplace.

Call it

curl -sS -X POST "$BASE/v1/pdf/markdown" \
  -H "X-API-Key: $DOCSMITH_KEY" \
  -F "file=@manual.pdf" -F "pages=1-20" | jq -r .markdown

curl -sS -X POST "$BASE/v1/pdf/text" \
  -H "X-API-Key: $DOCSMITH_KEY" \
  -F "file=@contract.pdf" \
  -F "pages=1-2"

curl -sS -X POST "$BASE/v1/pdf/images" \
  -H "X-API-Key: $DOCSMITH_KEY" \
  -F "file=@deck.pdf" -F "dpi=150" -F "format=png" -o page.zip

curl -sS -X POST "$BASE/v1/image/convert" \
  -H "X-API-Key: $DOCSMITH_KEY" \
  -F "file=@photo.jpg" -F "format=webp" -F "max_width=1600" -o photo.webp

What "Markdown" means here

Structure is recovered from the layout, not from a language model, so it is the same output every time and it costs the same per page as the raw text. The rules:

Heading levels come from relative font size and weight, and only sizes at least 8% apart get their own level.
Paragraphs are rejoined across visual line breaks, including hyphenated and soft-hyphenated words, using the document's own measured line spacing.
Bullet glyphs become - , numbered markers are kept as written, and --- marks each page boundary.
Standalone page numbers are dropped as furniture; body text beginning with Markdown syntax is escaped so it cannot forge a heading.
When the layout is ambiguous it stays a plain paragraph, because a false heading splits a chunk in the wrong place and nobody notices.

Pricing

plan	requests / month	price	notes
Free	100	$0	Same endpoints, 20MB uploads.
Starter	5,000	$9	Email support, 99% target.
Pro	50,000	$39	Highest render dpi, priority queue.

Calls that fail validation are not counted. A successful response carries X-Quota-Remaining and X-Quota-Reset.

What it does not do

No OCR. Scanned pages come back empty, and pretending otherwise would waste your time. /v1/pdf/markdown says so with a 422 no_text_layer instead of returning an empty document.
No table reconstruction and no reflowing of multi-column pages. Cells and columns arrive in reading order as text, which is right for search and wrong for spreadsheets - use the extracted text if you need the grid back.
No URL fetching. You upload the bytes; the server never makes an outbound request.
No storage. Files live in a per-request temp directory and are deleted before the reply is written.
No LLM calls, so latency is a function of page count, not of someone else's queue.