Capsule Architecture

A capsule is Rohrpost’s fundamental unit of compressed context. It captures the essential information from a conversation or workflow session in a structured, queryable format optimized for LLM consumption.

Capsule structure

Each capsule contains:

Component Purpose
Summary High-level overview of what was discussed or decided
Entities Key names, identifiers, and references extracted from context
Decisions Explicit decisions made during the session
Code refs File paths, function names, and architectural patterns mentioned
Embeddings Vector representations for semantic retrieval

Compression pipeline

The distillation pipeline transforms raw conversation context into a capsule through several stages:

  1. Chunking — Split the input into semantically coherent segments
  2. Extraction — Identify entities, decisions, and code references using heuristic and LLM-based extractors
  3. Summarization — Generate a concise summary preserving critical information
  4. Embedding — Compute vector embeddings for similarity-based recall
  5. Archival — Store the capsule with full provenance metadata

Token economics

A typical 8,000-token conversation compresses to a capsule of 1,500–2,500 tokens — a 65–80% reduction. When multiple capsules are recalled for a query, Rohrpost further ranks and truncates to fit within your configured token budget.

Multi-tenant isolation

Capsules are scoped to a tenant and project. The storage layer enforces strict isolation — a query in one tenant’s scope will never surface capsules from another tenant, even when sharing the same underlying database.