Using GPT-Vision for Drafting Patent Figure Descriptions

To use the GPT tool for figure description drafting, perform the following steps:

1. Open the GPT/LLM patent drafting tool

   From the ClaimMaster tab, click on the Patent Drafting Tools menu, then click on Generate Text with GPT/LLM menu:

   

2. Select one of available GPT/LLM templates for drafting a patent figure

   Pick one of the GPT prompt templates that ask GPT to write a description of the figure
   

   When populating the prompt template, ClaimMaster will automatically attempt to extract the following information from the figures/specification and fill-in these placeholders in the template to improve GPT output accuracy:
   [FIGURE_NUMBER] – specifies the figure #. 
   [FIGURE_SHORT_DESC] – specifies short description of the figure that would be extracted from the specification, if available.  
   [FIGURE_PARTS] – specifies the list of part #s extracted from the figures or  part names + numbers extracted from the Specification.

3. Select the file with one or more figures

   Select the file with figures in the middle section of the GPT tool by clicking on the magnifying glass button. ClaimMaster can process multi-figure files in Visio, PDF, PPT, and Word formats. In addition, you can also point the tool to single figure files in PNG or JPEG formats. If Annotate part #s in loaded figures checkbox is selected, ClaimMaster will insert red triangles next to part #s in the figures. This may help GPT better identify parts in the figures. However, if you have very dense figures with many parts closely located to each other, uncheck this box to avoid accidentally overwriting parts #s in the figures with annotations.

   

4. Select one of the figures from the file (Optional)

   If you’ve selected a file with multiple pages (e.g., PDF, Visio, PPT), then ClaimMaster will open a selection window allowing you to pick the specific page from the file for processing. You can preview the figure image on the right side of the window (if you click on the figure preview, you’ll be able to zoom in). Once you are done, click on the Select page button to pick one page for processing. ClaimMaster sends one page to GPT at a time. However, you can always pick another page from the same file later by clicking on the Change selected page button in the GPT tool.

   

5. Review and update the prompt that will be sent to the configured GPT service

   Once you pick the page from the file and a prompt template, you’ll see the preview of the prompt template that would be send to GPT. ClaimMaster will attempt to populate the prompt template with the information extracted from the figure and also fill-in short description of the figure and part #s, if available:
   [FIGURE_NUMBER] – specifies the figure #. 
   [FIGURE_SHORT_DESC] – specifies short description of the figure that would be extracted from the specification, if available.  
   [FIGURE_PARTS] – specifies the list of part #s extracted from the figures or  part names + numbers extracted from the Specification.
   If data is not available for these replacement fields, they will be removed from the template. You have two ways to add the list of part names and numbers to the prompt.

   First, you select a figure from which ClaimMaster can extract part #s, it will populate the GPT prompt (i.e., the [FIGURE_PARTS] field) with a list of part #s and you can add corresponding part names for each number directly in the prompt.
   

   In addition, ClaimMaster can automatically extract a list of part names + numbers from the open Word document to populate the prompt. To pull parts from the document, switch to the “Parts’ tab in the variables section, check the “Show with part numbers” checkbox and select one or more parts from the table to insert into the prompt (see video at 5:56 mark for how this works)
   

6. Send the prompt to the configured GPT service

   Once you review the prompt and are satisfied with how it looks, click on the Send prompt to LLM button to send both the figure and the corresponding prompt to the configured GPT service. The response from GPT usually takes anywhere from 10 to 40 seconds and sometimes you may need to repeat, if the GPT service times out. Note that image processing is only available for GPT vision models, so you’ll need to configure ClaimMaster to use the appropriate GPT model from OpenAI or Azure to enable this capability (e.g., gpt-4-vision-preview).

   

7. Review the results and insert them into the document

   Once you receive the response from the GPT service, you can review it the output window. You can then either edit it directly in the window or insert it into the document and then continue editing there.