It is difficult to find a balance between accuracy, scalability and cost-effectiveness, especially when faced with scenarios where hundreds of millions of pages of documents need to be processed, and the cost is often prohibitive.<\/p>\n\n\n\n To start using Gemini 2.0 to process PDF documents, you first need to set up the environment and create an inference client. Here are the specific steps:<\/p>\n\n\n\n With the above steps, we have completed the base environment configuration and successfully uploaded the first PDF file for processing. It is worth noting that the Gemini File API allows up to 20GB of files to be stored per project, with a maximum of 2GB per file, and uploaded files are saved for 48 hours.<\/p>\n\n\n\n Gemini 2.0 a powerful feature is the ability to extract structured data from PDF files. Below we will show how to use the actual case Pydantic model with Gemini to achieve this feature.<\/p>\n\n\n\n First, we define a generic method to process PDF files and return structured data:<\/p>\n\n\n\n For the invoice class PDF, we can define the following model to extract the key information:<\/p>\n\n\n\n For forms containing handwritten content, we can similarly define specialized models:<\/p>\n\n\n\n Through the above example, we can see that Gemini 2.0 can accurately identify the text content in the PDF, including even handwritten text, and will be converted to a structured JSON data format, greatly simplifying the data extraction process.<\/p>\n\n\n\n In RAG (Retrieval Augmented Generation) systems, document chunking is a key step in addition to basic text extraction.Gemini 2.0 allows us to do OCR and semantic chunking in a single step.<\/p>\n\n\n\n Here is a tip word for converting PDF to Markdown and semantic chunking at the same time:<\/p>\n\n\n\n This approach recognizes the semantic boundaries of a document and generates more meaningful chunks of text, greatly improving the accuracy of subsequent retrieval. Compared with the traditional mechanical chunking based on the number of characters, semantic chunking is better able to maintain the coherence and integrity of the content.<\/p>\n\n\n\n For more complex scenarios, we can define nested Pydantic models to handle multiple levels of data:<\/p>\n\n\n\n Here are some best practices for improving efficiency and accuracy when processing PDF documents at scale:<\/p>\n\n\n\n For the need to deal with a large number of PDF scenarios, you can achieve batch processing to improve efficiency:<\/p>\n\n\n\n Selecting the right model variant for the actual requirements can significantly reduce costs:<\/p>\n\n\n\n The following is a comparison of the processing efficiency of the different models:<\/p>\n\n\n\n In production environments, it is critical to implement robust error handling mechanisms:<\/p>\n\n\n\n For PDFs that contain complex forms, you can use the following cue words to improve form recognition accuracy:<\/p>\n\n\n\n![\"\"](\"https:\/\/school.myshirtai.com\/wp-content\/uploads\/2025\/05\/Jietu20250213-233957@2x.jpg\")
<\/figure>\n\n\n\nConfiguring the Environment and Setting Up Gemini 2.0<\/h2>\n\n\n\n
Install the necessary libraries<\/h3>\n\n\n\n
%pip install \"google-genai>=1\"\n<\/code><\/pre><\/div>\n\n\n\nCreating Clients and Model Configurations<\/h3>\n\n\n\n
from google import genai\n\n# Create client\napi_key = \"YOUR_API_KEY\" # Replace with your API key.\nclient = genai.Client(api_key=api_key)\n\n# Define the model to be used\nmodel_id = \"gemini-2.0-flash\" # Also use \"gemini-2.0-flash-lite-preview-02-05\" or \"gemini-2.0-pro-exp-02-05\"\n<\/code><\/pre><\/div>\n\n\n\nUploading and processing PDF files<\/h3>\n\n\n\n
# Upload PDF file\ninvoice_pdf = client.files.upload(file=\"invoice.pdf\", config={'display_name': 'invoice'})\n\n# See how many tokens the file is converted to\nfile_size = client.models.count_tokens(model=model_id, contents=invoice_pdf)\nprint(f'File: {invoice_pdf.display_name} equals to {file_size.total_tokens} tokens')\n\n# Sample output: File: invoice equals to 821 tokens\n<\/code><\/pre><\/div>\n\n\n\nStructured PDF data extraction practice<\/h2>\n\n\n\n
Define generic data extraction methods<\/h3>\n\n\n\n
def extract_structured_data(file_path: str, model: BaseModel).\n # Uploading a file to the File API\n file = client.files.upload(file=file_path, config={'display_name': file_path.split('\/')[-1].split('.') [0]})\n\n # Generating a structured response using the Gemini API\n prompt = f \"Extract the structured data from the following PDF file\"\n response = client.models.generate_content(model=model_id,\n contents=[prompt, file], config={'response_mime_content\n \n 'response_schema': model})\n\n # transforms the response into a Pydantic model and returns it\n return response.parsed\n<\/code><\/pre><\/div>\n\n\n\nCase 1: Invoice data extraction<\/h3>\n\n\n\n
from pydantic import BaseModel, Field\n\nclass Item(BaseModel).\n description: str = Field(description=\"The description of the item\")\n quantity: float = Field(description=\"The Qty of the item\")\n gross_worth: float = Field(description=\"The gross worth of the item\")\n\nclass Invoice(BaseModel).\n \"\"\"Extract the invoice number, date and all list items with description, quantity and gross worth and the total gross worth.\"\"\"\"\n invoice_number: str = Field(description=\"The invoice number e.g. 1234567890\")\n date: str = Field(description=\"The date of the invoice e.g. 2024-01-01\")\n items: list[Item] = Field(description=\"The list of items with description, quantity and gross worth\")\n total_gross_worth: float = Field(description=\"The total gross worth of the invoice\")\n\n# Extract the data using this model\nresult = extract_structured_data(\"invoice.pdf\", Invoice)\n\n# Output results\nprint(f \"Extracted Invoice: {result.invoice_number} on {result.date} with total gross worth {result.total_gross_worth}\")\nfor item in result.items: print(f \"Item: {item_gross_worth}\")\n print(f \"Item: {item.description} with quantity {item.quantity} and gross worth {item.gross_worth}\")\n<\/code><\/pre><\/div>\n\n\n\n![\"\"](\"https:\/\/school.myshirtai.com\/wp-content\/uploads\/2025\/05\/image-54.png\")
<\/figure>\n\n\n\nCase 2: Form processing with handwritten content<\/h3>\n\n\n\n
class Form(BaseModel).\n \"\"\"Extract the form number, fiscal start date, fiscal end date, and the plan liabilities beginning of the year and end of the year.\"\"\"\"\n form_number: str = Field(description=\"The Form Number\")\n start_date: str = Field(description=\"Effective Date\")\n beginning_of_year: float = Field(description=\"The plan liabilities beginning of the year\")\n end_of_year: float = Field(description=\"The plan liabilities end of the year\")\n\n# Extract data\nresult = extract_structured_data(\"handwriting_form.pdf\", Form)\n\n# output results\nprint(f'Extracted Form Number: {result.form_number} with start date {result.start_date}. \\nPlan liabilities beginning of the year {result.beginning_of_year} and end of the year {result.end_of_year}')\n# Output Example: Extracted Form Number: CA530082 with start date 02\/05\/2022.\n# Plan liabilities beginning of the year 40000.0 and end of the year 55000.0\n<\/code><\/pre><\/div>\n\n\n\nAdvanced Applications: Document Chunking and Semantic Understanding<\/h2>\n\n\n\n
PDF semantic chunking example<\/h3>\n\n\n\n
CHUNKING_PROMPT = \"\"\"OCR the following page into Markdown. Tables should be formatted as HTML.\nDo not surround your output with triple backticks.\nChunk the document into sections of roughly 250 - 1000 words. Our goal is\nOur goal is to identify parts of the page with same semantic theme.\nbe embedded and used in a RAG pipeline.\nSurround the chunks with html tags.\"\"\"\"\n\n# is processed using this hint word\nresponse = client.models.generate_content(\n model=model_id,\n contents=[CHUNKING_PROMPT, pdf_file], .\n)\n\nchunked_content = response.text\n<\/code><\/pre><\/div>\n\n\n\nComplex Data Extraction with Pydantic<\/h3>\n\n\n\n
class Person(BaseModel): first_name: str = Field(description=\"The first name of the person\")\n first_name: str = Field(description=\"The first name of the person\")\n last_name: str = Field(description=\"The last name of the person\")\n last_name: str = Field(description=\"The last name of the person\") last_name: str = Field(description=\"The last name of the person\")\n work_topics: list[Topic] = Field(description=\"The fields of interest of the person, if not provided please return an empty list\")\n\n# Generate a response using the Person model\nprompt = \"Philipp Schmid is a Senior AI Developer Relations Engineer at Google DeepMind working on Gemini, Gemma with the mission to help every developer to build and benefit from AI in a responsible way.\"\nresponse = client.models.generate_content(\n model=model_id,\n contents=prompt,\n config={'response_mime_type': 'application\/json', 'response_schema': Person}\n)\n\nThe # SDK automatically converts the response to a Pydantic model\nphilipp: Person = response.parsed\nprint(f \"First name is {philipp.first_name}\")\n<\/code><\/pre><\/div>\n\n\n\nPerformance Optimization and Best Practices<\/h2>\n\n\n\n
Batch Processing and Token Optimization<\/h3>\n\n\n\n
async def batch_process_pdfs(file_paths, model, batch_size=10):: results = [].\n results = []\n for i in range(0, len(file_paths), batch_size):: batch = file_paths[i:i+batch_size].\n batch = file_paths[i:i+batch_size]\n tasks = [extract_structured_data(path, model) for path in batch]\n batch_results = await asyncio.gather(*tasks)\n results.extend(batch_results)\n print(f \"Processed batch {i\/\/batch_size + 1}\/{(len(file_paths)+batch_size-1)\/\/batch_size}\")\n return results\n<\/code><\/pre><\/div>\n\n\n\nModel Selection and Cost Control<\/h3>\n\n\n\n
\n
mould<\/strong><\/th> PDF pages processed per dollar (Markdown conversion)<\/strong><\/th><\/tr><\/thead> Gemini 2.0 Flash<\/td> Approx. 6,000 pages<\/td><\/tr> Gemini 2.0 Flash Lite<\/td> Approx. 12,000 pages<\/td><\/tr> Gemini 1.5 Flash<\/td> Approx. 10,000 pages<\/td><\/tr> OpenAI 4-mini<\/td> Approx. 450 pages<\/td><\/tr> OpenAI 4o<\/td> About 200 pages<\/td><\/tr> Anthropic Claude-3.5<\/td> Approx. 100 pages<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n Error handling and retry mechanisms<\/h3>\n\n\n\n
def extract_with_retry(file_path, model, max_retries=3):: for attempt in range(max_retries)\n for attempt in range(max_retries).\n try.\n return extract_structured_data(file_path, model)\n except Exception as e: if attempt == max_retries\n if attempt == max_retries - 1: print(f \"Failed to get to the file.\n print(f \"Failed to process {file_path} after {max_retries} attempts: {e}\")\n return None\n print(f \"Attempt {attempt+1} failed, retrying: {e}\")\n time.sleep(2 ** attempt) # Exponential Retreat Strategy\n<\/code><\/pre><\/div>\n\n\n\n![\"\"](\"https:\/\/school.myshirtai.com\/wp-content\/uploads\/2025\/05\/rd-bench-example.jpg\")
<\/figure>\n\n\n\nForms Processing Optimization<\/h3>\n\n\n\n