PdfTranscriber

An agentic object that transcribes PDF pages into text and image descriptions using OCR and LLM vision.

Description

PdfTranscriber is an agentic object addressed to the user. The user calls transcribe() with a path to a PDF file and receives structured output containing text transcriptions and image descriptions for each page.

The transcriber produces high-quality results by combining two text sources:

1. LLM vision — text transcription and image description from a multimodal LLM.
2. Tesseract OCR — text recognition via tesseract.

When cross-referencing is enabled, both sources are unified to produce the most complete and accurate result.

Usage

transcriber = PdfTranscriber()
results = await transcriber.transcribe(
    "document.pdf",
    cross_reference=True
)
for page_num, (text, description) in enumerate(results, start=1):
    print(f"--- Page {page_num} ---")
    print(f"Text: {text}")
    print(f"Description: {description}")

transcribe(pdf_path: str, cross_reference: bool = False) -> list[tuple[str, str]]

Transcribe a multi-page PDF.

Args:

• pdf_path — Path to the PDF file.
• cross_reference — When True, cross-reference tesseract and LLM output per page. When False, use tesseract only (no LLM needed).

Returns: A list of (page_text, image_description) pairs, one per page.

Process:

1. Splits the PDF into single-page files using mutool.
2. Converts each page to a PNG image.
3. Runs tesseract OCR for text transcription.
4. If cross-referencing is enabled, runs the LLM vision agent for transcription and image description, then cross-references both sources.
5. If cross-referencing is disabled, uses the tesseract output directly.

Requirements:

• mutool (from mupdf) for PDF splitting and rendering.
• tesseract-ocr for OCR.
• An LLM backend configured via ChatBotManager (required only when cross_reference=True).

Internal Agentic Behavior

When cross_reference=True, the transcriber internally invokes its own agent twice per page:

• Once for text transcription and image description using the page image.
• Once for cross-referencing tesseract and LLM outputs into a unified result.

These internal invocations use persistent_thread_id=None (transient).