Improve routing function

Some ideas:

• Priority setting
  • speed:
    • no image -> STATIC_PARSE
    • image -> LLM_PARSE
      • Can also check if image size <<< page size and if so, default to static
  • accuracy: The current setting
    • has hidden hyperlink AND no image (or very small image) -> STATIC_PARSE
    • All other cases -> LLM_PARSE
    • The intuition behind the above routing is that LLM_PARSE (with a strong model) will usually represent the output much more accurately than STATIC_PARSE. However, the LLMs are currently not able to extract hidden (embedded) links from PDFs. STATIC_PARSE is also unable to process images (currently). So, if there are hidden links AND there are no images, we can route the doc to STATIC_PARSE.
• Check for tables in doc with STATIC_PARSE and reroute to LLM_PARSE if tables are present.
  • Reasoning: The LLM would be better at accurately representing the table - It can also use HTML to represent more complex tables (e.g.: those with merged cells).
• Custom routing functions:
  • An option can be provided to define a custom routing function by specifying the conditions for STATIC_PARSE vs LLM_PARSE.
  • One way this could be done is to pass a dictionary with keys for certain utility functions (e.g.: has_image_in_pdf, has_hyperlink_in_pdf) and boolean values to indicate whether they should be True or False in order for the document to be routed to STATIC_PARSE.

Start by locating the routing function and the existing STATIC_PARSE and LLM_PARSE paths. Review how images, hidden hyperlinks, tables, and priority or accuracy settings are currently detected; done means the routing behavior and any custom routing option match the specified conditions and are covered by tests.