Updates December 8
PGrid
- Added brand filter for SSD table
- There was an issue where ClickHouse went down and then the frontend showed empty tables. I had a previous fix for this where query errors should have stopped the build but this didn’t work. Updated build pipeline to have a ClickHouse health check before build and before upload to Cloudflare and fail build if health check fails.
- Updated classification pipeline stale item check to include CPU pipeline
- Updates to classification pipeline to standardise it more across CPU and SSDs
- Added some internal CLI commands for CPU classification similar to ones already there for SSDs
- Ran full CPU classification manually of ~10k items in queue for classification.
- Managed to set up trial Google Cloud account and use Google Vertex Kimi K2 thinking through OpenRouter through their BYOK so I didn’t have to deal with weird Google auth.
- Updated specs queue table to support deduping. Added a script for CPU models that passes batches of titles and gets it to come up with a CPU model that is used to dedupe the CPU specs search queue and could be extended to other categories. Running this reduced CPU model specs search queue from ~1500 to ~150
- Updated CPU description generation to now use BAML, be much simpler, output a specific scannable format for a CPU. Ran for all CPUs in DB. Updated frontend to display descriptions.