Improve the first part of the workflow

Currently the worklow extracts **relevant data** from the pdf reports by naming each page and searching for relevant topics(via semantic search). This approach is not very effective most of the times

Therefore add new functions in the api layer to:
- Fetch the index page of the pdf file
- Send the page to a multimodal llm to extract a json of {topic: page_num,...}
- Use this json and replace the existing semantic search logic with simple page number find of relevant topics

Note: You could use grok inference for free multimodal llms :)
PS: ensure your new functions are created in seprate js files and import necessary functions in routes file. Dont create entire logic in the routes file