Show HN: ParseHawk – 100% Local Document AI with API, CLI, and Web UI

Local-first document AI. Run 100% locally by default, with API, CLI, and Web UI.

Quickstart · First extraction · API, CLI, and Web UI · Requirements · Development

ParseHawk turns PDFs, scans, images, text files, and Markdown into structured JSON without sending sensitive documents to a third-party AI API. It is built for developers and teams working with private data: invoices, receipts, contracts, internal documents, customer files, medical or financial records, and other unstructured inputs that should stay under your control.

The default setup runs fully locally. ParseHawk uses vLLM on Linux NVIDIA machines and vLLM Metal on macOS Apple Silicon, so you can run practical document extraction on a server or even on your MacBook. You can drive the same workflow from the browser, from curl

, or from the parsehawk

CLI.

• Extract structured JSON from unstructured PDFs, scans, images, text, and Markdown
• Define your own schemas for the data you want back
• Run zero-shot extraction with only instructions and a schema
• Add few-shot examples when a document type needs more guidance
• Improve extraction quality without training a model
• Improve extractors over time with better instructions, schemas, and examples
• Get validated JSON output using JSON Schema Draft 2020-12
• Keep files, jobs, extractors, and results local by default
• Use the Web UI for humans and the REST API or CLI for scripts, services, and agents
• Control both the local stack and the extraction API from one parsehawk

CLI - Run on Linux with vLLM or on macOS Apple Silicon with vLLM Metal

ParseHawk runs on macOS Apple Silicon and Linux x86_64 with an NVIDIA GPU. Windows is not supported yet.

macOS Apple Silicon details #

Required:

uv

• Docker Desktop
• Xcode Command Line Tools
• Apple Silicon Mac with enough unified memory for NuExtract3-W4A16

Verified:

• MacBook Pro M3 Pro with 18 GB unified memory
• MacBook Pro M3 Pro with 36 GB unified memory

Recommended:

• 16 GB unified memory minimum for the default local workflow
• 32 GB or more for larger context lengths

Linux NVIDIA details #

Required:

uv

• Docker Engine
• Docker Compose
• NVIDIA driver
• NVIDIA Container Toolkit
• NVIDIA GPU with enough VRAM for NuExtract3-W4A16

Verified:

• NVIDIA L4 with 24 GB VRAM

Recommended:

• 16 GB VRAM minimum for the default local workflow
• 24 GB VRAM or more for larger context lengths

Run ParseHawk from a Git checkout with uv and install the CLI as an editable local tool:

git clone https://github.com/parsehawk/parsehawk.git
cd parsehawk
uv tool install --editable .
parsehawk start

Then open:

• Web UI: http://127.0.0.1:5173 - API docs: http://127.0.0.1:8000/docs - OpenAPI JSON: http://127.0.0.1:8000/openapi.json

Stop ParseHawk:

parsehawk stop

Check your local setup:

parsehawk doctor

The easiest first run is image-to-JSON extraction with the bundled receipt image and the seeded prebuilt Receipt

extractor.

• Start ParseHawk with parsehawk start

. - Open http://127.0.0.1:5173. - Upload .tests/fixtures/receipt/receipt.jpg

• Select the prebuilt Receipt

extractor. - Select the uploaded file and click Run extraction. - Inspect the extracted fields and JSON result.

Expected fields include:

{
  "merchant_name": "PARSEHAWK COFFEE",
  "receipt_id": "R-1001",
  "date": "2026-06-21",
  "total": 11.22,
  "currency": "EUR"
}
parsehawk files upload tests/fixtures/receipt/receipt.jpg
parsehawk extractors list
parsehawk extract \
  tests/fixtures/receipt/receipt.jpg \
  --extractor extractor_... \
  --wait

Use the Receipt

extractor ID from extractors list

.

API=http://127.0.0.1:8000

EXTRACTOR_ID=$(
  curl -s "$API/v1/extractors" |
    jq -r '.[] | select(.name=="Receipt" and .is_prebuilt==true) | .id'
)

FILE_ID=$(
  curl -s -X POST "$API/v1/files" \
    -F "upload=@tests/fixtures/receipt/receipt.jpg;type=image/jpeg" |
    jq -r '.id'
)

JOB_ID=$(
  curl -s -X POST "$API/v1/jobs" \
    -H "Content-Type: application/json" \
    -d "{\"extractor_id\":\"$EXTRACTOR_ID\",\"file_id\":\"$FILE_ID\"}" |
    jq -r '.id'
)

curl -s "$API/v1/jobs/$JOB_ID" | jq .

Jobs are asynchronous. Poll GET /v1/jobs/{job_id}

until status

is completed

or failed

.

ParseHawk exposes one local API. The CLI and Web UI are clients of that API. The CLI has two jobs: it controls the local ParseHawk stack (start

, stop

, status

, doctor

, restart

) and it works with the data plane (files

, extractors

, schemas

, jobs

, and one-shot extract

).

Core resources:

POST   /v1/files
GET    /v1/files
GET    /v1/files/{file_id}
GET    /v1/files/{file_id}/content
DELETE /v1/files/{file_id}

POST   /v1/schemas/validate

POST   /v1/extractors
GET    /v1/extractors
GET    /v1/extractors/{extractor_id}
PATCH  /v1/extractors/{extractor_id}
DELETE /v1/extractors/{extractor_id}

POST   /v1/jobs
GET    /v1/jobs
GET    /v1/jobs/{job_id}
DELETE /v1/jobs/{job_id}

Jobs return the canonical extracted JSON inline as job.result.data

once completed.

Useful CLI commands:

parsehawk files upload document.pdf
parsehawk files list
parsehawk schemas validate schema.json
parsehawk extractors create --name invoice_v1 --schema schema.json --instructions "Extract invoice fields."
parsehawk jobs create --extractor extractor_... --file-id file_...
parsehawk jobs get job_...
parsehawk extract document.pdf --schema schema.json --instructions "Extract invoice fields." --wait

Public IDs are TypeID-style strings with resource prefixes such as file_...

, extractor_...

, and job_...

.

An extractor combines:

• a name
• natural-language instructions
• JSON Schema Draft 2020-12
• optional few-shot examples
• optional thinking mode

A minimal extractor schema:

{
  "type": "object",
  "properties": {
    "invoice_number": {
      "type": ["string", "null"],
      "description": "The invoice number exactly as shown on the document."
    },
    "total_amount": {
      "type": ["number", "null"],
      "description": "The final total amount to pay."
    }
  },
  "required": ["invoice_number", "total_amount"],
  "additionalProperties": false
}

Few-shot examples can use inline text or uploaded files:

{
  "name": "invoice_v1",
  "instructions": "Extract the invoice fields exactly.",
  "schema": {
    "type": "object",
    "properties": {
      "invoice_number": { "type": ["string", "null"] }
    },
    "required": ["invoice_number"],
    "additionalProperties": false
  },
  "examples": [
    {
      "input": { "type": "text", "text": "Invoice #A-123" },
      "output": { "invoice_number": "A-123" }
    },
    {
      "input": { "type": "file", "file_id": "file_..." },
      "output": { "invoice_number": "B-456" }
    }
  ]
}

ParseHawk validates model output against the schema and stores the canonical result under job.result.data

.

The schema dialect is documented in docs/schemas/parsehawk-extraction-schema.schema.json. It supports JSON Schema plus optional

x-parsehawk.semantic

metadata for NuExtract3-oriented scalar hints.The default model is:

numind/NuExtract3-W4A16

ParseHawk talks to the runtime through an OpenAI-compatible API. On macOS, the runtime runs on the host through vLLM Metal because Metal acceleration is not available inside a normal Linux container. On Linux, the runtime runs as part of Docker Compose.

Current defaults:

Setting	Default
vLLM package	vllm==0.23.0
Linux runtime image	vllm/vllm-openai:v0.23.0
Model	numind/NuExtract3-W4A16
GPU memory utilization	0.5
Max model length	8192 by default, 32768 on larger Apple Silicon Macs
PDF render DPI	170
PDF max pages	25

Common overrides:

PARSEHAWK_VLLM_MAX_MODEL_LEN=16384 parsehawk start
PARSEHAWK_VLLM_GPU_MEMORY_UTILIZATION=0.6 parsehawk start
PARSEHAWK_VLLM_MODEL=numind/NuExtract3-W4A16 parsehawk start
PARSEHAWK_VLLM_IMAGE=vllm/vllm-openai:v0.23.0 parsehawk start

ParseHawk uses Pydantic settings. Common environment variables:

Environment variable	Default	Description
PARSEHAWK_DATA_DIR
data
Local storage directory for SQLite, uploaded files, logs, and local state.
PARSEHAWK_DATABASE_PATH
data/parsehawk.db
SQLite database path.
PARSEHAWK_LOG_LEVEL
INFO
Log level for API, worker, runtime, and Web UI logs.
PARSEHAWK_LOG_MODEL_IO
false
When true and PARSEHAWK_LOG_LEVEL=DEBUG , log model-runtime request and response JSON from the API/worker process. Image data URLs are redacted.
PARSEHAWK_INFERENCE_ENGINE
none
API/worker inference engine. parsehawk start sets this to vllm when a runtime is configured.
PARSEHAWK_VLLM_BASE_URL
http://127.0.0.1:8080/v1
OpenAI-compatible model runtime URL.
PARSEHAWK_VLLM_MODEL
numind/NuExtract3-W4A16
Model name sent to the runtime.
PARSEHAWK_VLLM_MAX_MODEL_LEN
platform-specific	vLLM context length. Overrides the automatic local default.
PARSEHAWK_VLLM_MAX_NUM_SEQS
128
Linux vLLM maximum concurrent decode sequences.
PARSEHAWK_VLLM_GPU_MEMORY_UTILIZATION
0.5
vLLM memory reservation fraction.
PARSEHAWK_VLLM_IMAGE
vllm/vllm-openai:v0.23.0
Linux Docker runtime image.
PARSEHAWK_VLLM_CACHE_HOME
~/.cache/vllm
Linux host cache for vLLM compile artifacts.
PARSEHAWK_PDF_MAX_PAGES
25
Maximum PDF pages rendered for one extraction.
PARSEHAWK_PDF_RENDER_DPI
170
PDF page image render DPI.
PARSEHAWK_TELEMETRY_DISABLED
false
When truthy, disables anonymous usage analytics.

CLI config:

parsehawk config list
parsehawk config set log.level DEBUG
parsehawk restart

ParseHawk collects anonymous usage analytics. Two events are sent to PostHog:

install

— once per install, the first time you start ParseHawk.run_started

— each time a user starts an extraction run.

Each event carries only coarse, non-identifying data:

• a random per-install id stored in data/telemetry-id

• the input type ( file

ortext

, on runs) - the ParseHawk version and your operating system name

• an approximate location (country/region)

ParseHawk never sends file contents, file names, extractor instructions, schemas, or extracted data, and it never creates a personal profile from the per-install id. The first time you run parsehawk start

or parsehawk dev

, you will see a notice describing this.

To opt out, set either of these before starting ParseHawk:

export PARSEHAWK_TELEMETRY_DISABLED=1
export DO_NOT_TRACK=1

When ParseHawk runs in Docker, these variables are passed through to the API and worker containers automatically.

Maintainers can tag internal usage instead of dropping it:

export PARSEHAWK_TELEMETRY_INTERNAL=1

By default ParseHawk stores local state under data/

:

data/
  parsehawk.db
  files/
  logs/
  parsehawk-state.json
  telemetry-id

Stop ParseHawk before deleting data/

:

parsehawk stop
rm -rf data
parsehawk start

If data/

is deleted while ParseHawk is still running, old processes can keep serving from already-open SQLite handles. parsehawk start

refuses to start when target ports are already occupied without a live state file. In that case, stop the process using the port and start again.

Development requires:

git

just

uv

pnpm

Useful commands:

just setup          # install dependencies and pre-commit hooks
just start          # product-like Docker mode
just dev            # local-source development mode
just web-dev        # Web UI dev server only
just smoke          # local smoke flow
just test           # Python tests
just e2e            # local end-to-end API tests (needs the model runtime up)
just format         # format Python
just lint           # Ruff linting
just typecheck      # ty type checking
just web-typecheck  # TypeScript checks
just web-test       # Web UI tests
just web-build      # production Web UI build
just check          # all standard checks
just hooks-run      # run pre-commit on all files

Pre-commit hooks are not installed automatically by Git. Run this once per clone:

just setup

The hooks run Ruff, ty, Python tests, Web UI typecheck, and Web UI tests. CI should still run the same checks; hooks are just the fast local feedback loop.

Development mode:

parsehawk dev

Product-like local mode:

parsehawk start

Start with the built-in health checks:

Check status:

parsehawk status

Read logs:

ls data/logs
tail -f data/logs/api.log
tail -f data/logs/worker.log
tail -f data/logs/runtime.log

Restart:

parsehawk restart

If Docker or the runtime gets into a strange state, stop ParseHawk before removing local data:

parsehawk stop
rm -rf data
parsehawk start

If the Model Runtime is slow to become ready, give it a few minutes on first startup while vLLM loads model weights, profiles memory, and warms kernels.

To start only the API and Web UI without local inference:

parsehawk start --runtime none

ParseHawk stands on excellent open-source projects, including:

FastAPIfor the API framework and OpenAPI docsvLLMandvLLM Metalfor local model servingNuExtract3for the default extraction modelPydantic,Ruff, anduvfor the Python toolchainReact,Vite, andTailwind CSSfor the Web UI

Near-term focus:

• make the macOS and Linux runtime paths boringly reliable
• publish an installable CLI package
• improve the Web UI schema builder
• add stronger end-to-end runtime smoke tests
• document deployment options for VPS and container platforms

Later:

• Python SDK
• migrations and PostgreSQL support
• batch extraction
• review/correction workflows
• eval tooling
• bring-your-own OpenAI-compatible runtime

ParseHawk is developed by Totoy GmbH in Vienna, Austria. If you are interested in an enterprise deployment, private-cloud setup, or managed infrastructure for sensitive document workflows, contact support@totoy.ai.

ParseHawk follows SemVer.

Until v1.0.0

, ParseHawk is in developer preview. Breaking changes may happen in any minor release, for example from v0.1.0

to v0.2.0

.

Patch releases, such as v0.1.1

, are intended to be backward-compatible bug fixes for that minor line.

We will move to v1.0.0

once the core CLI commands, the REST API, and the config file format are stable enough for users to rely on.

ParseHawk is open source under the Apache-2.0 license. See LICENSE.

Third-party dependencies retain their own licenses.

source & further reading

github.com — original article