Document OCR

POST /api/document-ocr5 quota units per call · cache hits free

Run OCR on a document or photo and return the recognised text + word/line bounding boxes. Works on PNG, JPEG, WebP, AVIF.

Extract text from images with Tesseract's LSTM engine (English language model bundled). Handles document scans, screenshots, and reasonably-clean photos. Inputs: - `image`: PNG, JPEG, WebP, or AVIF bytes (the front Worker auto-converts non-PNG inputs via the codecs before feeding them to ). - `granularity`: how much detail to return. `text` (default, just the concatenated string), `word` (text + per-word boxes + confidence), `line` (text + per-line boxes + confidence), or `word+line` (everything). Response modes (negotiated via Accept header): - `application/json` (default): `{text, durationMs, width, height}` plus `words[]` and/or `lines[]` when granularity requests them. Each box is `{rect: {left, top, right, bottom}, text, confidence}`. - `text/plain`: just the recognised text as a UTF-8 body — same content as `text` in the JSON envelope. Boxes are discarded in this mode (they only fit in JSON). For best results pre-process noisy scans with `document-prepare-for-ocr` first (deskew + binarize). Phase B of `docs/EXPANSION_PLAN.md`. Other languages can be added later by lazy-loading additional `.traineddata` files from R2 — currently only English is bundled.

Inputs

Name	Type	Default	Description
image*	file	—	Input document image (PNG/JPEG/WebP/AVIF).
granularity	enum (text \| word \| line \| word+line)	"text"	Level of detail to return (boxes are only emitted in JSON mode).

Response

Modes: json, text. Cache: yes (24h TTL).

Every response carries x-cache (HIT / MISS / BYPASS), x-cache-signature (stable across identical inputs), and the rate-limit headers listed below.

Quota & limits

Cost per call	5 units against the 10,000-unit monthly quota. Cache hits are free — only cache misses decrement.
Burst limit	200 requests per rolling minute, shared across all tools.
Max request size	`25 MiB` (base64 inflates file payloads ~33% — the limit applies to the decoded bytes).
Max response size	`5 MiB`
Timeout	60s wall clock.
Rate-limit headers	`X-RateLimit-Limit`, `X-RateLimit-Remaining`, `X-RateLimit-Burst-Limit` on every response; `X-Quota-Warning` once 80% of the monthly quota is spent; `Retry-After` on 429s.

Errors

Errors are JSON with an error string; schema failures add a details object (Zod's flattened field errors).

Status	When	Body
400	Input fails the tool's schema (wrong type, missing required field, malformed base64) or the file bytes can't be parsed as the expected format.	{ "error": "invalid inputs", "details": { "fieldErrors": { … } } }
401	Missing or invalid X-Api-Key / X-Account-Id headers (or no session when called from the browser).	{ "error": "unauthorized" }
404	Unknown tool id, or calling /api/* on the website hostname instead of api.xtract.bot.	{ "error": "not found" }
413	Request body over the tool's max request size, or an image header declaring more megapixels than the tool's pixel budget.	{ "error": "…exceeds maxRequestBytes…" }
415	Content-Type isn't application/json on the json-body transport.	{ "error": "unsupported content-type; expected application/json" }
429	Monthly quota or the 200/min burst limit exhausted. Check Retry-After and the X-RateLimit-* headers.	{ "error": "monthly quota exceeded", "monthRemaining": 0, … }
5xx	Conversion engine failure. Safe to retry; if it persists, the input is hitting a bug — please report it.	{ "error": "…" }

Code samples

Built from the invoice example.

# Download or substitute the example input:
#   curl -O https://xtract.bot/examples/ocr-fixtures/document-small.png
IMAGE=$(base64 -w0 < document-small.png)

curl -X POST https://api.xtract.bot/api/document-ocr \
  -H "Content-Type: application/json" \
  -H "Accept: application/json" \
  -H "X-Account-Id: $XTRACT_ACCOUNT_ID" \
  -H "X-Api-Key: $XTRACT_API_KEY" \
  -d '{
  "granularity": "word",
  "image": "'"$IMAGE"'"
}'