Kiah Box Projects & Build Notes

The Problem

Large Language Models process tokens sequentially, making long prompts expensive and slow. Current solutions either lose context (summarization) or maintain full size (no optimization).

The Solution

SPEC was an experimental prompt pre-processor designed to reduce prompt tokens and latency while avoiding breakage on structured content. It is preserved for inspection, but the measured production benefit did not justify a meaning-changing Kiah preprocessing layer.

• Token-Focused Compaction - removes low-information filler and stopwords (lossy by design)
• Protected Segments - leaves code blocks, inline code, URLs, emails, paths, and quoted strings untouched
• Profiles - safe, balanced, aggressive trade safety vs savings
• Shorthand Codec - optional lossless encode/decode via ~-tokens (for storage/transport)
• Go/No-Go Benchmark - uses Ollama prompt_eval_count + latency to measure real token impact

Results

Example

Status

Technical Details

Disposition

The code remains available as an archived research artifact, with these operating boundaries:

• Do not enable it in Kiah or production prompt paths
• Do not treat character reduction as proof of token or latency value
• Compactor is not reversible (prompt-only); codec is lossless for dictionary tokens
• Any future compression work starts as a new measured candidate