https://www.boundaryml.com logo
Powered by
🟡 Losing integer IDs in merging; seeking tips for...
# questions
r
I’m losing a lot of integer IDs in merging blocks of 100, any tips other than shrinking the block size? Eventually, NVIDIA needs to show 800 source IDs… that’s too many. How do I address this context length problem? I may have to break it up and paginate or something? Or convert it to edge data during the extraction? Would that help?
a
is this a new regression or has it always been like this?
you could try adding 'reasoning' text fields in between each filled in ID
will be way more tokens but it's something to explore
r
It’s been there but started acting up. Claude Code suggested a specific instruction to always return all ids and it fixed it 🙂
What’s a reasoning text field?
Why they merged?
a
Copy code
{
  property1: string
  reasoning: string
}
r
I mean, what are its semantics?
And why would that help?
a
it's just adding thoughts during the generation, which can help it reason about what the next field should be
or about what is missing for example
r
Latency?
a
yep it would definitely add latency, since it's just gonna generate more tokens for each reasoning field
r
Ahhh yes 0.216.0 made a couple timeout tasks pass.
@aaronv (BAML) do you have any recommendations to users affected by the sunsetting of Gemini 2.0 Flash? 2.5 Flash Lite costs the same but performs poorly.
a
unfortunately not at the moment, currently not super familiar with these two models. Whats the performance issues you see?
r
Flash Lite drops integer IDs when merging records
2 Views
PreviousNext
Powered by