Table of Contents

Interface IContextCompressor<T>

Namespace
AiDotNet.Interfaces
Assembly
AiDotNet.dll

Defines the contract for compressing context documents to reduce token usage while preserving relevance.

public interface IContextCompressor<T>

Type Parameters

T

The numeric data type used for relevance scoring.

Remarks

A context compressor reduces the size of retrieved documents before passing them to a language model, helping to stay within token limits while maintaining the most relevant information.

For Beginners: A context compressor is like a summarizer for search results.

Think of it like preparing a briefing:

  • You retrieve 10 long documents (might be 50,000 tokens)
  • Your language model can only handle 8,000 tokens
  • The compressor extracts key sentences and information
  • Result: Compressed to 5,000 tokens with the most important content

This ensures you can:

  • Use more retrieved documents without hitting token limits
  • Focus on the most relevant parts of each document
  • Reduce costs by using fewer tokens
  • Still get accurate answers from the compressed context

Methods

Compress(List<Document<T>>, string, Dictionary<string, object>?)

Compresses a collection of documents while preserving relevance to the query.

List<Document<T>> Compress(List<Document<T>> documents, string query, Dictionary<string, object>? options = null)

Parameters

documents List<Document<T>>

The documents to compress.

query string

The query text used to determine relevance.

options Dictionary<string, object>

Optional compression parameters.

Returns

List<Document<T>>

The compressed documents with reduced content but maintained relevance.

Remarks

This method takes documents and reduces their content size while keeping the parts most relevant to the query. The compression strategy varies by implementation - some may use extractive methods (selecting key sentences), others may use abstractive methods (generating summaries), or hybrid approaches.

For Beginners: This reduces document size while keeping important information.

For example:

  • Input: 3 documents, each 2000 tokens (6000 total)
  • Query: "How do I train a neural network?"
  • Processing: Extract sentences about training, remove unrelated content
  • Output: 3 documents, each 500 tokens (1500 total)

The compressed documents still contain the key information needed to answer the question, but take up much less space.

Edit this page