Docroom Document Compression Engine

Problem

Large tool outputs, JSON, logs, and search results can consume context quickly, but replacing them with summaries can hide fields, versions, anomalies, and evidence locations. docroom targets large, reconstructable, navigable office materials and provides a compress-and-retrieve workflow instead of an unauditable summary.

How It Works

• Classify before compressing: A–E responsibility levels decide eligibility. Only structured, repetitive, retrievable A/B tool outputs proceed; C/D/E source facts, Office/HTML structure, and sensitive material remain intact.
• Use the official Headroom MCP: calls headroom_compress, headroom_retrieve, and headroom_stats for native compression, CCR original-text cache, and token accounting.
• Fidelity gates: preserve source tool, field names, URLs, IDs, versions, pagination boundaries, anomalies, and evidence locations before compression; trim only duplicate templates, irrelevant fields, and exact duplicates.
• Validate benefit: requires tokens_saved > 0, recoverable metadata, and consistent CCR retrieval; otherwise falls back to pagination, narrower reads, or keeping the original.

Boundaries

Best for large, repetitive, structured tool outputs that will be reused later. Not suitable for exact citations, final conclusions, native Word/Excel/PPT structure, images, charts, formulas, DOM/CSS, sensitive material, or unauthorized content. The runtime must expose the official Headroom MCP, and factual answers should be verified by headroom_retrieve(hash).