DeepSeek-OCR-WebUI

DeepSeek-OCR-WebUI

Self-hosted DeepSeek-OCR-2 for documents and images — vLLM or transformers on NVIDIA, native Apple Silicon, and CPU, with per-page quality checks.

🌐 English · 简体中文 · 繁體中文 · 日本語

What's new · Screenshots · Quick start · API · Architecture · Performance · Development

---

✨ What's new in v5.0

v5 is a ground-up renovation of the backend and a hardening of the Vue 3 frontend. Every number below was measured; see CHANGELOG for the full list.

	v4.1	v5.0
Apple Silicon / CPU	loads, then fails on first page (upstream code hard-codes .cuda())	works — device-agnostic inference, M5 Max 8.1 pages/min
NVIDIA engines	transformers only, one page at a time	vLLM (continuous batching, 132 pages/min on an A10G) or transformers
Old GPUs (T4 / V100, no bf16)	BFloat16 vs float crash (#30 #40)	fp16 + eager, verified on T4
Driver requirement	CUDA 13 image ⇒ driver ≥ 580 (#46)	CUDA 12 builds
Docker image	13 GB with weights inside (#38)	weights in a volume; hf image ≈ 4.0 GB unpacked
Output quality	HTTP 200 = "success"	per-page ok + warnings (repetition, truncation, coordinate loops, empty)
Readiness	/health 200 before the model can read	503 until a warm-up page is read back correctly
Large PDFs	200 pages ⇒ 662 MB in IndexedDB, freezes (#31 #50)	78 MB, parsed once in a worker; 500 MB upload limit (#51)
Markdown export	dropped multi-box blocks (#49)	100 % character recall on all samples
Security	binds 0.0.0.0, CORS * with credentials, XSS via v-html	127.0.0.1 by default, CORS allow-list, DOMPurify
Tests	0 backend tests; 123 frontend tests failing on Node 26	backend, frontend, e2e and mutation suites — live counts in the badges above

---

📸 Screenshots

All screenshots are generated by frontend/scripts/screenshots.mjs against a real backend (DeepSeek-OCR-2 on an Apple M5 Max). The demo document is generated by scripts/make_demo_docs.py.

Document → Markdown with layout boxes and a rendered table

Batch queue	Second PDF page
Recognition modes	Service status (engine · device · precision)
Swagger UI	Mobile (简体中文)

---

🚀 Quick start

🔒 The service has no login. Every command below listens on 127.0.0.1 only. To use it from another machine, use an SSH tunnel (ssh -L 8001:127.0.0.1:8001 you@gpu-host) or put it behind a reverse proxy that authenticates users — do not publish the port on all interfaces.

Option 1 — Docker, NVIDIA GPU (vLLM engine, recommended)

git clone https://github.com/neosun100/DeepSeek-OCR-WebUI.git && cd DeepSeek-OCR-WebUI
docker compose up -d                  # builds the image; first start downloads ~6.8 GB of weights
open http://127.0.0.1:8001            # UI ·  API docs: http://127.0.0.1:8001/docs

Prebuilt images (published by CI on each release):

docker run -d --gpus all --shm-size 8g -p 127.0.0.1:8001:8001 -v dsocr-models:/models \
  ghcr.io/neosun100/deepseek-ocr-webui:5.0.0          # vLLM engine
docker run -d --gpus all --shm-size 8g -p 127.0.0.1:8001:8001 -v dsocr-models:/models \
  ghcr.io/neosun100/deepseek-ocr-webui:5.0.0-hf       # transformers engine (smaller, also for T4/V100)

/health returns 503 while the model is loading (first start: download + load + warm-up) and 200 once a warm-up page has been read back correctly. docker ps shows healthy at that point.

Option 2 — macOS (Apple Silicon) or Linux without Docker

Requires uv and Node.js ≥ 20.

git clone https://github.com/neosun100/DeepSeek-OCR-WebUI.git && cd DeepSeek-OCR-WebUI
(cd frontend && npm ci && npm run build)       # the backend serves the built UI
uv sync --extra hf                             # transformers + torch (MPS on macOS, CUDA 12.6 wheels on Linux)
uv run deepseek-ocr-webui serve                # http://127.0.0.1:8001

On Linux with vLLM instead: uv sync --extra vllm && uv run deepseek-ocr-webui serve --engine vllm.

Offline / air-gapped

Download once (huggingface-cli download deepseek-ai/DeepSeek-OCR-2 --revision aaa02f3811945a91062062994c5c4a3f4c0af2b0 --local-dir /models/ocr2), then start with DSOCR_MODEL_PATH=/models/ocr2.

Configuration

All settings are environment variables with the DSOCR_ prefix (config.py is the single source). The most useful ones:

Variable	Default	Meaning
DSOCR_HOST / DSOCR_PORT	127.0.0.1 / 8001	Listen address. The Docker image sets 0.0.0.0 inside the container; the published port is the boundary.
DSOCR_ENGINE	auto	auto · vllm · hf (auto = vLLM if installed and CUDA is available)
DSOCR_DEVICE	auto	transformers engine: cuda · mps · cpu
DSOCR_REQUIRE_GPU	false	fail fast instead of silently falling back to CPU
DSOCR_VLLM_URLS	—	comma-separated external vLLM servers (multi-GPU / multi-host, least in-flight)
DSOCR_MODEL_PATH	—	local weights directory (no network)
DSOCR_MAX_UPLOAD_MB · DSOCR_MAX_PDF_PAGES	200 · 500	upload limits
DSOCR_MAX_QUEUE · DSOCR_MAX_PER_CLIENT · DSOCR_MAX_PER_IP	32 · 1 · 4	admission control (HTTP 429 when exceeded)
DSOCR_CORS_ORIGINS	—	allow-listed origins (off by default)
DSOCR_TRUST_PROXY	false	trust X-Forwarded-For (only behind your own proxy)

v4 variable names (PORT, LOCAL_MODEL_PATH, FORCE_BACKEND) still work.

---

🧩 Recognition modes

Mode	Prompt	Output
document (default)	`<	grounding
ocr	`<	grounding
free	Free OCR.	plain text, fastest
figure	Parse the figure.	chart / figure description
describe	Describe this image in detail.	natural-language description
find	document parsing + server-side block search	the layout blocks that contain find_term (block-level, see below)
freeform	your prompt	anything

About find: DeepSeek-OCR-2 has no native "locate this phrase" prompt — 12 out of 12 trials with the v1-style Locate <|ref|>…<|/ref|> prompt fell back to document parsing. v5 therefore parses the layout and returns the boxes of the blocks whose text contains your term. Precision is per block, not per word.

---

🔌 API

Interactive docs: /docs (Swagger) and /redoc. The OpenAPI spec is checked in at docs/openapi.json and kept in sync by a test.

# One image
curl -s -F [email protected] -F prompt_type=document http://127.0.0.1:8001/api/ocr | jq '{ok, warnings, text}'

# A PDF, streamed page by page (NDJSON: start → page* → done)
curl -sN -F [email protected] -F pages=1-3 http://127.0.0.1:8001/api/ocr/pdf

# Status / capabilities
curl -s http://127.0.0.1:8001/health | jq '{status, ready, engine, warmup}'
curl -s http://127.0.0.1:8001/api/capabilities | jq '.limits, .known_weak'

Each page result contains text, raw_text (with <|ref|>/<|det|> grounding tags), boxes (pixel coordinates), finish_reason, ok and warnings:

Warning	Meaning	Fails ok?
repetition / repetition_cut	the model looped at the end; the loop is trimmed (and stopped while generating)	yes
truncated	hit the generation limit	yes
coordinate_loop	box coordinates degenerated ([23,268,268,268…])	yes
empty_output	nothing was generated	yes
near_empty	≤ 20 letters/digits — usually a blank page	no
bbox_invalid	some boxes were out of range and dropped; text kept	no
find_no_match	find mode: no block contains the term	no

Errors always look like {"success": false, "detail": {"error", "message", "actual", "limit"}, "request_id"}. The v4 endpoints /ocr, /pdf-to-images and the v4 /health fields keep working.

MCP (Claude, Cursor, …)

{ "mcpServers": { "deepseek-ocr": {
    "command": "uv", "args": ["run", "--directory", "/path/to/DeepSeek-OCR-WebUI", "--extra", "mcp", "deepseek-ocr-webui-mcp"],
    "env": { "DSOCR_BASE": "http://127.0.0.1:8001" } } } }

Tools: ocr_image(path | image_base64, mode, find_term, custom_prompt, grounding), ocr_pdf(path, mode, pages), health().

---

🏗 Architecture

• Browser — PDFs are rendered in a Web Worker (each file parsed once, bytes transferred once), pages are kept in IndexedDB so long jobs survive reloads, and Markdown / DOCX / PDF are generated locally.
• Service (server/deepseek_ocr_webui) — admission control, upload sniffing and pixel limits, the engine, per-page judges, grounding parsing.
• Engines — vllm streams from vllm serve and closes the stream as soon as a repetition loop is detected; hf re-implements the upstream preprocessing so tensors follow the model's real device and dtype.

---

📊 Performance

Measured with scripts/bench.py on dedicated machines (12 pages: sample images, rendered PDF pages, synthetic pages with known text; document mode). Full record, memory, agreement and reproduction steps: docs/BENCHMARKS.md.

Hardware	Engine	Concurrency	Pages / min	p50 latency	Known-text recall
NVIDIA A10G 24 GB	vLLM 0.33	1	46.4	0.81 s	1.00
NVIDIA A10G 24 GB	vLLM 0.33	16	132.1	5.88 s	1.00
NVIDIA A10G 24 GB	transformers (bf16)	1	9.5	2.05 s	1.00
NVIDIA T4 16 GB	transformers (fp16)	1	7.8	2.6 s	1.00
Apple M5 Max	transformers (MPS, bf16)	1	8.1	3.3 s	1.00

vLLM and transformers agree on the same pages with a mean bidirectional character recall of 0.985 (min 0.849 on one dense page). transformers handles one page at a time, so extra concurrency only queues.

---

🛡 Limits and known weak spots

From GET /api/capabilities (known_weak lists the public benchmark evidence): DeepSeek-OCR-2 is weaker on photos with perspective / skew, handwriting, and low-resource languages. Defaults: 200 MB per upload, 80 MP per image, 500 PDF pages per request (rendered at 144 DPI), 8192 tokens per page.

---

🧪 Development

uv sync --extra hf --extra mcp --group dev
uv run pytest -n auto tests/unit tests/regression        # FakeEngine, no GPU needed
DSOCR_REAL_MODEL=1 uv run pytest tests/e2e/test_real_model.py   # real weights on CUDA / MPS
DSOCR_BASE=http://127.0.0.1:8001 uv run pytest tests/e2e         # against a running server
cd frontend && npm ci && npx vitest run && npx playwright test --project=chromium

• Mutation testing — tests/mutations.ini lists deliberate bugs (judges, admission, security defaults, readiness, engines, MCP); each must turn the suite red (run with touchstone mutate.py, command in the file header).
• GPU verification — scripts/gpu-verify/ builds both images on a GPU machine, runs the e2e suite against each, benchmarks them and compares their outputs.
• Generated assets — node frontend/scripts/screenshots.mjs, uv run python scripts/badges.py [--check], deepseek-ocr-webui openapi > docs/openapi.json.
• CI: GitHub Actions (backend · frontend · e2e · image build) and GitLab CI.

Upgrading from v4

• Same port (8001), same UI, /ocr and /health stay compatible. The Docker image no longer contains the weights: mount a volume at /models.
• Container listens on 0.0.0.0 inside, compose publishes on 127.0.0.1 only — change the publish deliberately if you need LAN access, and add authentication in front of it.
• find now requires find_term and returns block-level matches.

---

👥 Contributors

The Vue 3 frontend — including its E2E suite — was contributed by @cloudman6 and is the foundation v5 builds on.

📈 Star history

🤝 Contributing

Issues and pull requests are welcome. Please run the backend and frontend test suites before opening a PR; CI runs them on every pull request.

📄 License

MIT. DeepSeek-OCR-2 weights are distributed by DeepSeek under their own license.

🙏 Acknowledgments

DeepSeek-AI for DeepSeek-OCR-2 · vLLM for the native OCR-2 support and NGram logits processor · @cloudman6 and everyone who reported issues.