Updates June 15

PGrid

• work in progress for item enrichment
  • queued 100+ officeworks item enrichment tasks but only a small percentage saved data. investigated crawler and seems to be proxy issue. probably similar issue with bhphoto. started looking at alternative vpns with residential support
  • updated bhphoto single page scraper and tested it
  • added centrecom single page crawler
  • noticed item descriptions can be very long and useless for classification and specs search taking up space in the db. experimenting with using a single llm step for extra long descriptions to extract anything useful
    • ran evals for different providers for deepseek flash. some providers always return issue responses which baml doesn’t parse and some work much better. updated to only use 1 or 2 providers
• continuing work on specs search
  • minor fixes to specs queue deduping. now deletes keys even if llm generates no keys
  • ran claude opus to iterate on monitor specs search prompts after using ~40 existing model rows created by codex agent. found that model name definition was poor and there was mixing of random specs in the monitor names. fixed and iterated.
  • so now monitor model specs search has both codex agent + baml workflow support
  • added laptop style verification to monitor specs search
• updating on how classification and specs search works overall
  • added llm metrics tracking. this is to mainly help understand rate limits for vertex but think it will be useful in general. tracks tokens and errors for llm calls.
  • updating baml client selection to allow dynamic setting of existing clients like classification small llm client. this will allow easier switching of llms for me instead of me commenting out specific lines in the baml clients file
  • updating classification runs to have an option where if there is an llm error it will stop the run and clean up in progress rows. this is a configurable option
• added new skill to allow agents to query logs from production workers
• debugging local .git folder ballooning in size and making push_hosthatch2 command slow. due to codex adding turn step snapshots and not ignoring in progress pg backups. updated to ignore more