FormGuide

AI-assisted pre-check for Wohngeld applications. FormGuide converts uploaded income-proof PDFs to text with Docling (GPU), asks a local Gemma 3 27B model (via Ollama) whether the document is valid, and extracts applicant name, first name, and monthly income. A minimal Flask UI/API handles uploads and shows the result.

Features

• PDF-to-text conversion with Docling configured for CUDA 12
• LangChain agent using ChatOllama (Gemma 3 27B) for validation and data extraction
• Simple Flask frontend (/) plus JSON upload endpoint at /upload
• Sample documents in documents/ for quick smoke tests

Requirements

• Python 3.12+ and Pixi (or another env manager)
• NVIDIA GPU with CUDA 12 (Docling is initialized for GPU use)
• Ollama running locally with gemma3:27b available
• Ports: Flask defaults to http://127.0.0.1:5000; Ollama defaults to http://127.0.0.1:11434

Quickstart

# 1) Create/enter the env
pixi install
pixi shell

# 2) Start Ollama in another terminal (only once you may need to pull the model)
ollama pull gemma3:27b
ollama serve

# 3) Run the Flask app (inside the Pixi shell)
python formguide/app.py

# 4) Open the UI
# Visit http://127.0.0.1:5000 and upload a PDF from the documents/ folder.

How it works

1. DocumentProcessor (formguide/doc_process.py) converts the uploaded PDF to text using Docling with GPU acceleration.
2. FormGuideAgent (formguide/agent.py) is a LangChain agent backed by ChatOllama (Gemma 3 27B). It first checks if the document looks like an income proof (JA/NEIN), then extracts (Nachname, Vorname, Monatseinkommen) when valid.
3. run_analysis (formguide/processor.py) orchestrates conversion + agent calls and returns (success: bool, message: str).
4. app.py exposes /upload (multipart form field file) and renders the simple HTML interface at /. Uploaded files are saved under uploads/.

API response

POST /upload returns JSON: {"success": true|false, "error": "<text>"}. On success the error field currently carries the extracted income string; on failure it contains the rejection reason (frontend expects this format).

Repository layout

• formguide/app.py – Flask entrypoint and routes
• formguide/processor.py – pipeline wiring document conversion and agent logic
• formguide/doc_process.py – Docling PDF-to-text helper
• formguide/agent.py – LangChain + Ollama agent and prompts
• formguide/templates/index.html – minimal upload UI
• formguide/static/ – styling assets
• documents/ – example PDFs and extracted text for manual testing

Development tips

• Uploads are stored in uploads/; delete contents if you need a clean run.
• To exercise the pipeline without the UI, run python -m formguide.processor after placing a test file in uploads/ and adjusting the path in __main__.
• Ruff configuration is present in pyproject.toml; install and run it if you add linting to the workflow.***