Your Private Paperwork Assistant

A document assistant that searches your German legal documents and answers questions in English, with source citations. Built with Elasticsearch, Jina AI models on Elastic Inference Service, and a local LLM via Ollama.

Blog post: Your Private Paperwork Assistant

How it works

1. Embed — jina-embeddings-v5-omni-small on EIS embeds both PDF text and phone photos into the same vector space. A text query in English finds a German document without OCR.
2. Index — Vectors and extracted text go into Elasticsearch with a German analyzer for BM25.
3. Retrieve — Hybrid search combines BM25 keyword matching and kNN semantic search, merged with Reciprocal Rank Fusion (RRF).
4. Rerank — jina-reranker-v3 on EIS re-scores candidates using listwise cross-document attention.
5. Generate — A local LLM (Ollama) answers from the retrieved passages. No document text leaves your machine for this step.

Prerequisites

• Elastic Cloud deployment with EIS enabled (or self-managed Elasticsearch 9.x)
• Python 3.11+
• Ollama installed locally

Setup

# Clone and install
git clone https://github.com/sofia-vas/paperwork-assistant.git
cd paperwork-assistant
pip install -r requirements.txt

# Configure credentials
cp .env.example .env
# Edit .env with your Elastic Cloud ID and API key

# Pull a local model
ollama pull llama3.2:3b

# Generate the sample corpus (fictional German legal documents)
python generate_corpus.py

# Index documents into Elasticsearch via EIS
python ingest.py --force-recreate

# Start the chat interface
PORT=8000 python app.py

Open http://localhost:8000 in your browser.

Sample corpus

The repository includes a generator (generate_corpus.py) that creates 35 fictional German legal documents and 10 phone-photo versions:

• Employment contracts and payslips
• Health insurance notices and medical invoices
• Rent increase letters and utility bills
• Tax assessments and bank statements
• Telecom contracts and government notices

All documents are fictional (addresses in "Musterstadt", fictional IBANs) and safe to use for demos.

Architecture

User question (English)
        │
        ▼
┌─────────────────────────┐
│  Elasticsearch          │
│  ┌───────────────────┐  │
│  │ BM25 (German)     │──┤
│  │ kNN (image embed) │──┤── RRF merge ── Rerank (.jina-reranker-v3)
│  │ kNN (text embed)  │──┤
│  └───────────────────┘  │
└─────────────────────────┘
        │
        ▼ top passages
┌─────────────────────────┐
│  Ollama (local LLM)     │
│  Answer in English      │
│  with source citations  │
└─────────────────────────┘

Configuration

Variable	Default	Description
ELASTIC_CLOUD_ID	—	Elastic Cloud deployment ID
ELASTIC_API_KEY	—	Elasticsearch API key
ELASTICSEARCH_URL	—	Alternative: direct cluster URL
OLLAMA_MODEL	llama3.2:3b	Local LLM model name
OLLAMA_BASE_URL	http://localhost:11434	Ollama server URL
PORT	5000	Web server port

License

MIT