Skip to content

Using the LLM-based PDF Parser

  • Converts PDF content into Markdown format using Multimodal models.

  • Uses multimodal models to describe images within PDFs.

  • Supports page-wise or chunk-based processing for optimized performance.

Initializing the LLM-based PDF Parser

Make sure you have set up your API key for whichever model you specify in model_name below.

You can initialize the LLM PDF parser as follows:

parsing_config = ParsingConfig(
    n_neighbor_ids=2,
    pdf=PdfParsingConfig(
        library="llm-pdf-parser",
        llm_parser_config=LLMPdfParserConfig(
            model_name="gemini-2.0-flash",
            split_on_page=True,
            max_tokens=7000,
            requests_per_minute=5,
            timeout=60,  # increase this for large documents
        ),
    ),
)

Parameters

model_name

Specifies the model to use for PDF conversion. Default: gemini/gemini-2.0-flash

max_tokens

Limits the number of tokens in the input. The model's output limit is 8192 tokens.

  • Default: 7000 tokens (leaving room for generated captions)

  • Optional parameter

split_on_page

Determines whether to process the document page by page.

  • Default: True

  • If set to False, the parser will create chunks based on max_tokens while respecting page boundaries.

  • When False, the parser will send chunks containing multiple pages (e.g., [11,12,13,14,15]).

Advantages of False:

  • Reduces API calls to the LLM.

  • Lowers token usage since system prompts are not repeated per page.

Disadvantages of False:

  • You will not get per-page splitting but groups of pages as a single unit.

If your use case does not require strict page-by-page parsing, consider setting this to False.

requests_per_minute

Limits API request frequency to avoid rate limits.

  • If you encounter rate limits, set this to 1 or 2.