lamisema 1.1.0

Structured Information Extraction for Nepali and multilingual PDFs

LamiSema

Structured information extraction for Nepali PDFs.

---

The problem

Most Nepali PDFs silently return wrong output when you use standard tools on them. There are three types, each needing a different approach:

PDF type	Example source	Standard tool result
Unicode-native	Modern government portals	✅ Works fine
Legacy-encoded	Pre-2010 docs using Preeti/Kantipur font	❌ Returns garbage (g]kfn instead of नेपाल)
Scanned	Physical forms, old records	❌ Returns empty string

LamiSema detects the type first and automatically routes to the right strategy.

---

How it works

flowchart LR
    PDF[PDF Input] --> P[Pre-flight\nDetect encoding type]
    P -->|unicode_native| T[Text layer\npdfplumber]
    P -->|legacy_encoded| O[OCR\nTesseract nep+eng]
    P -->|scanned| O
    T --> N[NER + Date\nNormalization]
    O --> N
    N --> J[Structured JSON\nwith confidence scores]

---

Install

pip install lamisema

System dependency — Tesseract with the Nepali language pack:

# macOS
brew install tesseract tesseract-lang

# Ubuntu / Debian
sudo apt-get install tesseract-ocr tesseract-ocr-nep

---

Python usage

from lamisema import LamiSema

pipeline = LamiSema()

with open("report.pdf", "rb") as f:
    result = pipeline.extract(f.read(), filename="report.pdf")

print(result.encoding_type)        # "legacy_encoded"
print(result.overall_confidence)   # 0.74
print(result.pages[0].entities)    # [Entity(type="DATE_BS", text="२०८१ साल असार १५", ...)]

---

REST API

Start the server:

lamisema serve
# → http://localhost:9001/docs

# 1. Upload
curl -X POST http://localhost:9001/upload -F "file=@report.pdf"
# → { "doc_id": "DOC-A1B2C3D4" }

# 2. Detect encoding (fast, no extraction)
curl http://localhost:9001/preflight/DOC-A1B2C3D4

# 3. Extract everything
curl -X POST http://localhost:9001/extract/DOC-A1B2C3D4

# 4. Get result
curl http://localhost:9001/result/DOC-A1B2C3D4

Method	Endpoint	Description
GET	/	Health check
POST	/upload	Upload a PDF, get doc_id
GET	/preflight/{doc_id}	Encoding type + font analysis
POST	/extract/{doc_id}	Full extraction + NER
GET	/result/{doc_id}	Retrieve completed result
POST	/normalize-dates	Normalize BS dates in raw text

---

Try the demo app

A full-stack demo (Next.js frontend + API + MinIO) is in application-demo/.

cd application-demo
cp .env.example .env
docker compose -f docker-compose.local.yaml up --build
# → http://localhost:3000

---

Docs

Full documentation at lamisema.readthedocs.io

• Installation
• Python API
• REST API Reference
• Encoding types explained
• Contributing

---

License

MIT — see LICENSE