Mistral OCR API

FastAPI + Jinja2 templates · deployable to Google Cloud Run.

Uses mistral-ocr-latest (OCR 4.1).

Setup

cd ~/ocr-mistral-api
cp .env.example .env
# edit .env — set MISTRAL_API_KEY=...

That's it. Local dev loads .env automatically (python-dotenv). Cloud Run deploy passes the same file via --env-vars-file.

Local

pip install -r server/requirements.txt
cd server && uvicorn main:app --reload --port 8080

Cloud Run

chmod +x scripts/deploy-cloud-run.sh
./scripts/deploy-cloud-run.sh

Reads MISTRAL_API_KEY, MISTRAL_OCR_MODEL, and optional GEMINI_* from .env (strips PORT — Cloud Run sets that itself).

403 Forbidden in browser? Deploy succeeded but public access wasn't granted. A project Owner must run:

gcloud run services add-iam-policy-binding mistral-ocr-api \
  --region=europe-central2 --member=allUsers --role=roles/run.invoker

(Or grant you Cloud Run Admin / run.services.setIamPolicy.) Org policy may also block allUsers — check iam.allowedPolicyMemberDomains.

Build only (CI):

gcloud builds submit --config=cloudbuild.yaml
./scripts/deploy-cloud-run.sh

API

Method	Path	Description
GET	/	Upload UI
GET	/health	Readiness + model
POST	/api/ocr	OCR upload. After Mistral succeeds, Gemini fixes obvious typos (spellcheck in the JSON). Fail-soft if Gemini is down or unset. Form: spellcheck_enabled=true

Tests

pip install -r requirements-dev.txt
pytest -v

.env

Variable	Required	Default
MISTRAL_API_KEY	yes	—
MISTRAL_OCR_MODEL	no	mistral-ocr-latest
GEMINI_API_KEY	no	— (skips spellcheck if unset)
GEMINI_MODEL	no	gemini-2.5-flash
GEMINI_SPELLCHECK	no	1 (0 to disable)
PORT	local only	8080

Don't commit .env.