Hey folks, has anyone else noticed Codeium's inline suggestions seem to degrade in quality the longer/more complex your open file gets? I've been using it for about six months now, mostly for SQL and Python in my data pipeline work, and I'm hitting a weird pattern.
When I start a fresh `.sql` file—say, for a new fact table—the completions are sharp. It nails CTE structures, suggests relevant window functions, even gets the warehouse-specific syntax right (BigQuery vs Snowflake). But once the same file grows past maybe 200 lines with multiple complex joins and subqueries, the suggestions start feeling... off.
A couple examples from yesterday:
* In a long transformation script, it started suggesting columns that weren't in the joined tables.
* It repeated a `WHERE` clause that was already written 20 lines above.
* For a `CASE` statement, it hallucinated a non-existent `status` column from a different part of the codebase.
My hunch is the context window or the token limit might be prioritizing the *most recent* lines too heavily, losing the broader schema context. I'm curious if this is a known quirk or if I need to tweak something.
**My setup:**
- VSCode, Codeium extension v1.6.2
- Working with dbt projects, so lots of `.sql` and `.yml` files open
- Usually have 3-5 editor tabs open in the same window
Has anyone found workarounds? Splitting files into smaller modules helps, but sometimes you need a monolithic script for a complex transformation. Wondering if other BI/reporting folks are seeing this with their SQL or Python ETL scripts.
--diver
Data is the new oil - but it's usually crude.