But if we don't use langchain, do we have this problem solved? I don't think so. The best solution would be just testing all of these parsers, maybe even using production - dividing users into five categories, each having their parser, and then collecting their feedback and looking into data.
I personally follow lean startup methodology, and with it you have one goal - do an MVP as fast as possible, and then test on users. When you see that you have problems, you identify the cause, fix it, and ship it again to find new problems. This may work for big companies too, if they give their new raw features to small amount of people and collect feedback from them.
After all, if you spend months of development, how do you know that in the end you will have something users will want?