inklight

See documents. Understand meaning. Take action.
A local-first, OCR-powered document agent for Python — fast on a laptop CPU, and it shows its evidence.

Point Inklight at a folder of PDFs, scans or photos and ask for what you need, in plain language:

inklight ask ./invoices "Read all these invoices and create a CSV of vendor, date, tax, and total" --out ./out
inklight ask ./contracts "Find every contract that expires within 90 days"
inklight ask ./reports "Compare these reports and explain the differences"

You get an answer plus the page, line and bounding box it came from, a confidence score, warnings when something looks off, and a trace of every step the agent took.

Why Inklight

• Real OCR, on CPU. Built on Indra OCR: 43 ms per digital PDF page, 98 ms per scanned page on a laptop (benchmarks). Digital PDFs use their text layer first, and OCR runs only where needed.
• An agent, not a chatbot wrapper. It works out the intent, then runs a minimal plan of typed tools (search, extract, filter, calculate, compare, find inconsistencies, summarise, write). Every run has step limits, time budgets, retries and cancellation.
• Shows its work, and its limits. Every result says how your request was understood ("vendor contains Acme; total > 5,000 USD") and lists anything it could not apply, instead of quietly answering a broader question. Values come from labelled fields, extractive spans or quoted sentences with page and position. A citation proves where a value came from, not that it answers your question, so confidence is reported twice: understood (did it read your request right) and evidence (OCR, labels, subtotal + tax = total). Amounts in different currencies are never added or compared to each other.
• Small-model intelligence, measured. Optional models are each under 200M parameters: a 66M reader that can only point at text in your document, a 33M embedder for search, and an optional 135M writer whose output must be grounded. Calculations and filters are exact Python, not model guesses.
• Local and private by default. No API key needed. Nothing leaves your machine unless you configure a remote model.
• Safe with untrusted documents. "Ignore previous instructions" inside a PDF is treated as data and flagged. The agent has no shell, no code execution, and writes files only where you allow. See SECURITY.md.

Install

pip install inklight            # Python 3.11+, Windows / Linux / macOS, CPU only
inklight models pull            # optional: 66M reader + 33M embedder (~100 MB, once)

Indra's OCR models (~76 MB) download automatically on the first OCR run. After that, everything works offline.

60-second tour

inklight process ./documents/                                   # OCR, classify and index (incremental)
inklight ask ./documents/ "Find all invoices above $5,000"
inklight extract invoice.pdf --schema invoice.json --format csv
inklight serve                                                  # web workspace on http://127.0.0.1:8765
inklight watch ./inbox -r "extract vendor and total as csv" -o ./out   # automation

from inklight import DocumentAgent

agent = DocumentAgent()
result = agent.ask("Extract the invoice number, date, and total", files=["invoice.pdf"])

print(result.answer)        # human-readable answer
print(result.table())       # [{'document': 'invoice.pdf', 'invoice_number': 'INV-2041', 'date': '2026-10-03', 'total': 5616.0}]
print(result.sources[0])    # name='invoice.pdf' page=1 bbox=(200.4, 325.0, 667.6, 359.5) excerpt='INVOICE #INV-2041' ocr_conf=1.0
print(result.warnings, result.confidence, result.trace)

More: typed extraction without natural language (agent.extract(files, fields=[...]) or schema=), ask_async, library management (agent.library.search/remove/reindex), and the HTTP API at /api/docs when serving. See docs/API.md.

What it can do

How it works

request ─► planner (typed plan, from your words only) ─► bounded runner ─► tools ─► cited answer + what it understood
                                                                   │
            library: SQLite + FTS5 (+ vectors) ◄── ingest ◄── Indra OCR (text layer first, OCR where needed)

Details in docs/ARCHITECTURE.md. The short version of the "200M-parameter" story: we measured a 135M model as a planner and it was right 27.5% of the time, so the small models do what small models are good at (pointing at evidence, ranking passages) and deterministic code does the rest.

Configuration

~/.inklight/config.toml (see config.example.toml) or environment variables such as INKLIGHT_LIMITS__MAX_PAGES=50. Brain modes:

• rules: no models.
• auto (default): use the small models if they are installed.
• local: require the small models.
• models.remote_*: optional stronger model behind an OpenAI-compatible API, for low-confidence questions.

Limits, honestly

• OCR quality is Indra's. It is weaker on handwriting, degraded scans, complex tables and some scripts (Indra's notes). Low OCR confidence is surfaced on every value.
• Field labels are English. Ambiguous dates like 03/04/2026 are read month-first.
• The rule planner covers common document tasks. Unusual phrasing falls back to question answering, or to a remote model if you configured one.
• Semantic search scans vectors in memory, which suits libraries up to tens of thousands of chunks.

⚠ License notice — read before commercial use

Inklight's code is Apache-2.0. The Indra OCR model weights (downloaded by indra-ocr, not bundled) use a modified OpenRAIL-M license. It is free for research, personal use, and organisations under $20M revenue and under $20M funding. It restricts commercial use that competes with Datalab. A commercial document-AI product built on Inklight + Indra needs a license from Datalab, or a different OCR engine (the OCR layer is pluggable). Full audit: THIRD_PARTY_LICENSES.md.

Development

pip install -e ".[dev]"
pytest                      # 128 tests, about 1 minute on CPU (downloads Indra models on first run)
ruff check src tests && mypy src
python bench/run.py         # reproduces docs/BENCHMARKS.md

See CONTRIBUTING.md, CHANGELOG.md, CODE_OF_CONDUCT.md and, for maintainers, docs/PUBLISHING.md.