Context Compression and Summarization Protocols

Compression Necessity

Context often exceeds model context windows. Compression and summarization enable using large context volumes while respecting token limits. ECM Protocol defines standard approaches.

Compression Operations

Compression Request

{
  "operation": "context.compress",
  "source_contexts": ["ctx-1", "ctx-2", "ctx-3"],
  "target_tokens": 4000,
  "strategy": "hierarchical_summary",
  "preserve": {
    "entity_mentions": true,
    "temporal_markers": true,
    "key_facts": ["customer_tier", "issue_history"]
  }
}

Compression Response

{
  "compressed_context": {
    "content": "...",
    "token_count": 3850,
    "compression_ratio": 0.15
  },
  "source_mapping": [
    {"source": "ctx-1", "contribution": 0.4},
    {"source": "ctx-2", "contribution": 0.35},
    {"source": "ctx-3", "contribution": 0.25}
  ]
}

Compression Strategies

Extractive Summarization

Select most relevant sentences/passages. Preserve original wording. Good for factual content. Lower compression ratios.

Abstractive Summarization

Generate new text capturing key information. Higher compression possible. May introduce inaccuracies. Better for narrative content.

Hierarchical Compression

Multi-level compression for large contexts. Summarize chunks, then summarize summaries. Preserves structure at appropriate granularity. Efficient for very large contexts.

Selective Preservation

Keep specific information verbatim. Compress surrounding context. Critical facts preserved exactly. Configurable preservation rules.

Quality Metrics

Information Retention

Measure preserved information. Entity preservation rate. Fact accuracy after compression. Semantic similarity to original.

Compression Ratio

Track compression achieved. Tokens before/after. Trade-off compression vs. retention.

Protocol Extensions

Compression-specific extensions:

x-ecm-compression-strategy: Strategy used
x-ecm-compression-ratio: Achieved ratio
x-ecm-original-tokens: Pre-compression token count
x-ecm-preservation-score: Information retention metric

Caching Compressed Context

Cache compressed versions for efficiency. Key by source contexts + target tokens. Invalidate when sources change. Pre-compute common compressions.

Conclusion

Context compression enables large context usage within model limits. The ECM Protocol defines standard compression operations, strategies, and quality metrics for interoperable compression implementations.

> Context Compression and Summarization Protocols