semantic-summarization

# Semantic Summarization (Context Compression) ## Goals - Reduce token count by 70-90% - Preserve all entities, numeric values, relationships, and action items - Remove redundancy, verbosity, and low-information content ## Compression strategy (tiered) ### Tier 1 — Lossless (50-70% compression) Keep all facts. Remove adjectives, adverbs, filler, meta-commentary. ``` Input: "The function appears to be intentionally designed to handle edge cases like null values and undefined inputs gracefully." Output: "Function handles edge cases: null, undefined." ``` ### Tier 2 — Semantic (70-85% compression) Merge related facts. Abstract examples into patterns. Drop tangential details. ``` Input: "We tried using approach A which failed, then approach B which also failed, and finally approach C which worked." Output: "A, B failed. C succeeded." ``` ### Tier 3 — Extractive (85-95% compression) Keep only: entities, decisions, blockers, action items, key numbers. Drop: explanations, context, rationale, alternatives not chosen. ## Rules - Numbers, dates, percentages, IDs → always preserve verbatim - Proper nouns, filenames, function names → always preserve - Action items and decisions → always preserve - Repeated information → keep first occurrence only - Code snippets → keep signature + 1-line summary; drop body - Error messages → keep error type + code; drop full trace - Logs → keep timestamps + severity + message; drop metadata ## Output format - Use bullet points (not prose) - Prefix: `[K]` = key fact, `[D]` = decision, `[A]` = action, `[B]` = blocker