Contextual Compression
Trimming the fat. Why send 1000 tokens when 50 will do?
1. The Retrieval Paradox
You retrieve 10 documents because you want High Recall. But now your context window is full of irrelevant noise. If the answer is in the 7th document, the LLM might miss it ("Lost in the Middle").
2. The Compressor
A Contextual Compressor sits between the Retriever and the LLM. It takes the 10 retrieved docs and asks: "Given the user query, what parts of these docs are actually useful?"
It acts like a highlighter pen.
3. Techniques
A. LLMChainExtractor
Uses a small, fast LLM to rewrite the chunk.
- Original: "The CEO, John Smith, announced earnings today. He also mentioned he likes golf. Revenue is up 5%."
- Query: "What is the revenue?"
- Compressed: "Revenue is up 5%."
B. EmbeddingsFilter
Embeds the query and the sentences within the chunk. Only keeps sentences with high similarity. Cheaper than LLM extraction.
4. Summary
Retrieval gets the document. Compression extracts the answer. This creates a higher quality "Context Density" for the final generation model.
Key Intuition: "Don't pay for tokens you don't need."