Can I Upload PDFs Through the OpenAI API for Conversational Use?

Below is an overview of the current situation regarding PDF uploads through the API:

Current API File Upload Support

• The existing file upload endpoint (POST https://api.openai.com/v1/files) is designed primarily for tasks such as fine-tuning or batch processing.
• For fine-tuning and batch operations, the API requires files in specific formats (typically .jsonl), and there are file size and format restrictions in place.
• While the Assistants API mentions support for files of specific types and token limits, it does not list PDFs as a supported file type.

ChatGPT UI vs. API Functionality

• The ChatGPT UI for premium users offers a “chat with PDF” feature. However, that functionality is implemented internally in the ChatGPT interface and is not currently exposed as part of the public API.
• There is no public API endpoint that accepts PDF files and processes them in a way similar to the ChatGPT UI.

Recommendations for Your CLI Tool

• Preprocess the PDF: You may consider using a PDF extraction library (such as pdfminer for Python) to convert the content of your math texts into plain text.

• Chunking and Context Management: Once you have extracted the text, you could break it into manageable chunks (keeping token limits in mind), then feed those chunks to an OpenAI model via the chat or completion endpoints.

• Post-Processing: Process the model’s outputs to combine or refine answers as needed.

This approach allows you to leverage the powerful language processing capabilities of OpenAI’s models even without direct PDF upload support via the API.

Example: Extracting Text from a PDF (Python)

Below is a simple example using the PyMuPDF library to extract text from a PDF:

PYTHONCopied
import fitz  # PyMuPDF

def extract_text_from_pdf(pdf_path):
    doc = fitz.open(pdf_path)
    text = ""
    for page in doc:
        text += page.get_text()
    return text

pdf_text = extract_text_from_pdf("your_file.pdf")
print(pdf_text)

After extracting the text, you can then send it to the API using the chat or completions endpoints.

Final Note

At this time, while you cannot directly upload a PDF via the API for "chatting with" it, converting the PDF to plain text is a viable workaround for integrating math texts into your CLI tool.