I really liked this stuff you guys did

PDFs with tables are a univseral problem 😅 Sadly no one has it quite figured out yet, but it's on our list to improve.

Unstructured has some basic table stuff, we recently did a demo here
https://docs.llamaindex.ai/en/stable/examples/query_engine/sec_tables/tesla_10q_table.html#joint-tabular-semantic-qa-over-tesla-10q

Curious what you mean by "junk" with the sentence window -- probably because the data isn't actually easily split into sentences?

Technical documents often have lots of short sentences that don't make sense out of context, so embedding those is just going to add noise to the db

That's a fair point -- the auto-merging retriever is another similar approach that seems to also work well in my experience, that might make more sense for your data

As for tables, I built an index manually from pdf files by converting them to docx using Word itself, and then writing some code (with the help of AI, of course) to convert docx to markdown, with markdown-formatted tables. I then used a markdown parser/splitter from langchain to keep sections together, and to gather section meta-data, and then finally converted those docs to llama index Documents. This makes the tables readable by humans, since they are well-formed markdown, and hopefully readable by the AI too.

Did you have to use the UI in word, or was there a way to achieve this using purely code?

Would be cool if llama-index had a pipeline that did something similar

I didn't find an immediate way to do it in code, but I also didn't need to for my use case.