Gemini CLI sparks local spreadsheet extraction debate

A practical answer is usually not “LLM only” but “use deterministic spreadsheet tooling first, then let a model handle the ambiguous parts.” Local models can fit well when privacy matters or extraction rules are stable, but they usually need more plumbing and tuning than a hosted CLI. For well-structured spreadsheets, `pandas`, `duckdb`, or spreadsheet-specific scripting will beat any model on reliability and repeatability. Gemini CLI makes sense if you want quick multimodal or long-context assistance without building a custom pipeline. Local models are better when the data cannot leave the machine, but quality drops fast on messy tables, OCR noise, or inconsistent sheet layouts. The highest-leverage setup is usually hybrid: parse and normalize with code, then use an LLM for classification, field mapping, and exception handling. For scale, the bottleneck is often I/O and schema cleanup, not raw model intelligence.