Improved PDF Processing

We've improved how we process PDFs to better handle visual content.

Until now, when you uploaded a PDF to Langdock, we extracted the text from it and sent it to the AI model, and created so-called “embeddings” to enable the model to perform a semantic search on the PDF. This allows the model to generate qualitative responses based on the uploaded file.

To further improve this processing, we now also take screenshots of pages and provide them to the model. In simple terms, this gives the model “eyes”, so you can now ask questions like “Describe the details of the graphic on page 26" and get accurate answers, since the model can now actually "see" the document content. The model now has a new “view page” tool for inspecting pages, which is particularly helpful for images, charts, logos, and other visual elements.

As always, let us know what you think! 🙌

  • File size support increased: We increased the character limit for sizes from 4 million to 8 million characters.
  • Project limit increased: We increased the limit on the number of projects that can be created from ten to 30.
  • Voice input limit increased: We increased the limit of the voice input duration from five to ten minutes.
  • Performance improvements: Our engineering team has deployed larger performance improvements over the last few weeks, resulting in faster page loading times and smoother answer generation.
  • Model key pooling: Bring-your-own-key customers can now pool model deployments in separate regions. This helps increase rate limits and allows for another region to serve as a fallback in case the model encounters an error.
  • Updated documentation: Over the last year, more features have been added to the product, and our documentation has grown accordingly. We cleaned up the structure, translated the documentation into German, and launched documentation for workflows. You can find the documentation here: https://docs.langdock.com/