What You’ll Build
In about two hours, you will have an AI form-filling system that reads source documents such as passports, invoices, contracts, or medical records, extracts relevant fields, and maps them to target form fields automatically. Upload a stack of supporting documents and a blank form template, and the system returns a pre-filled form ready for human review. Processing 50 forms takes under 10 minutes on a single dedicated GPU server.
Insurance applications, government filings, immigration paperwork, and patient intake forms demand repetitive copying of data between documents and form fields. Errors from manual transcription create downstream delays and compliance risks. An AI form filler on GPU hardware eliminates transcription errors while keeping sensitive personal data entirely on your own infrastructure using open-source models.
Architecture Overview
The system has three stages: document OCR and extraction using PaddleOCR for scanned inputs, intelligent field mapping powered by an LLM via vLLM, and form population using template-based output generation. Source documents pass through OCR and document AI to extract text, tables, and structural elements. The LLM then maps extracted data to form field labels using semantic understanding rather than rigid rules.
LangChain orchestrates the pipeline with a two-pass approach. The first pass extracts all available data points from source documents into a structured profile. The second pass takes each form field, queries the profile for the matching data, and fills it in with appropriate formatting. A RAG layer can optionally pull reference data like postal codes, country codes, or industry classifications from a lookup database.
GPU Requirements
| Workload | Recommended GPU | VRAM | Forms Per Hour |
|---|---|---|---|
| Simple forms (10-20 fields) | RTX 5090 | 24 GB | ~120 forms/hr |
| Complex forms (50+ fields) | RTX 6000 Pro | 40 GB | ~80 forms/hr |
| Multi-document extraction | RTX 6000 Pro 96 GB | 80 GB | ~200 forms/hr |
OCR and LLM run sequentially per form, with OCR using GPU acceleration for fast image processing. The LLM handles the intelligent mapping step. An 8B model manages simple field mapping effectively; complex multi-document cross-referencing benefits from larger models. See our self-hosted LLM guide for model sizing.
Step-by-Step Build
Deploy PaddleOCR and vLLM on your GPU server. Define your form templates as JSON schemas listing field names, types, and validation rules. Build the extraction pipeline that processes source documents and creates structured data profiles.
# Form filling pipeline
EXTRACT_PROMPT = """Extract all personal and document data from this text.
Return a JSON object with every identifiable field:
name, date_of_birth, address, passport_number, etc.
Document text:
{ocr_text}"""
FILL_PROMPT = """Map the extracted data to form fields.
Extracted data: {extracted_profile}
Form fields requiring values:
{form_field_list}
Return JSON mapping each field name to its value.
Format dates as DD/MM/YYYY. Use title case for names.
If a field cannot be filled from available data, set it to null."""The output module takes the filled JSON and renders it into the target format, whether that is a PDF overlay, a web form submission, or a structured data export. Add a review interface where operators verify pre-filled values before final submission. Follow our vLLM production guide for optimal batch processing configuration.
Performance and Accuracy
On an RTX 6000 Pro, processing a multi-page source document through OCR takes 1.5 seconds per page. Field extraction and mapping for a 30-field form completes in 2 seconds. End-to-end processing of a typical insurance application with three source documents and a 40-field form takes about 12 seconds. Field-level accuracy exceeds 96% for clearly printed documents and 89% for handwritten inputs.
Validation rules catch formatting errors and impossible values before output. A confidence score per field helps reviewers focus attention on low-confidence entries rather than checking every field. The system learns from corrections when you feed reviewed forms back as training examples for prompt refinement.
Deploy Your Form Filler
Automated form filling eliminates hours of manual data entry per day in document-heavy industries. Insurance, immigration, healthcare, and legal operations see immediate ROI from reduced processing time and error rates. Deploy on GigaGPU dedicated GPU hosting to keep all personal data on-premises while processing forms at machine speed. Explore more build patterns in our use case library.
