Getting Started

Install and use forgetless in your project.

Installation

toml

[dependencies]
forgetless = "0.1"

Basic Usage

rust

use forgetless::{Forgetless, Config};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let result = Forgetless::new()
        .config(Config::default().context_limit(10_000))
        .add(&large_content)
        .run()
        .await?;

    println!("{}", result.content);
    Ok(())
}

output

Input:  247,000 tokens
Output: 9,842 tokens

Config Options

rust

use forgetless::{Forgetless, Config};

let result = Forgetless::new()
    .config(Config::default()
        .context_limit(128_000)  // Max output tokens
        .vision_llm(true)        // LLM for image descriptions
        .context_llm(true)       // LLM for smart scoring
        .chunk_size(256)         // Target chunk size
        .parallel(true)          // Parallel processing
        .cache(true))            // Embedding cache
    .add_file("diagram.png")
    .add_file("research.pdf")
    .query("Explain the architecture")
    .run()
    .await?;

Priority

rust

use forgetless::{Forgetless, Config, WithPriority};

let result = Forgetless::new()
    .config(Config::default().context_limit(50_000))
    .add(WithPriority::critical("System prompt"))  // Always kept
    .add(WithPriority::high(&conversation))        // High priority
    .add(&documents)                               // Medium (default)
    .add(WithPriority::low(&logs))                 // Low priority
    .run()
    .await?;

Files

rust

use forgetless::{Forgetless, Config, FileWithPriority};

let result = Forgetless::new()
    .config(Config::default().context_limit(100_000))
    .add_file("README.md")
    .add_file(FileWithPriority::high("main.rs"))
    .add_files(&["lib.rs", "config.rs"])
    .run()
    .await?;

Result

rust

result.content         // Optimized content
result.total_tokens    // Output token count
result.stats.input_tokens       // Input token count
result.stats.compression_ratio  // e.g., 14.5
result.stats.chunks_processed   // Total chunks
result.stats.chunks_selected    // Selected chunks

How It Works

Content is split into semantic chunks
Each chunk is embedded locally (all-MiniLM-L6-v2)
Chunks are scored using hybrid algorithm:
- Algorithmic: priority, recency, position
- Semantic: similarity to query (if provided)
- LLM: intelligent scoring (if enabled)
Top chunks selected within token budget
Chunks assembled into final optimized context