This is the absolute core of ContextFlow. Compressors programmatically strip redundant token pollution before calling an LLM, saving immense API costs and dropping Time-To-First-Token (TTFT) significantly.
Unlike traditional RAG systems that ask an LLM to "summarize this context," ContextFlow uses pure Python execution.
- AST Protection: It maps regular expressions over the
ContextItemtext to physically remove and safely cache Fenced Code Blocks (```python) and Agent Function Tool JSON blobs ({ "action": "xyz" }). - Redundancy Scrubbing: Once the fragile schema blocks are stored independently, the compressor iterates through the volatile text lines: removing whitespace, blank carriage returns, and completely ignoring cyclical linear duplicates commonly found in massive error crash-logs or LangGraph loop traces.
- Reassembly: The clean AST schema blocks are instantly rewritten perfectly identically into the compressed volatile text.
For absolute extreme edge cases (where files exceed 100,000 tokens), an alternate compressor exists which uses a fast auxiliary LLM (like gpt-4o-mini) to compress massive strings densely before inserting them into an expensive Claude 3.5 Sonnet or GPT-4o reasoning graph.