Configuration

`Config` covers the runtime switches most projects need. Lower-level tuning is available through `ForgetlessConfig`, `ChunkConfig`, `ScoringConfig`, and `TokenizerModel`.

Runtime options

1use forgetless::{Config, Forgetless};
2
3let result = Forgetless::new()
4    .config(
5        Config::default()
6            .context_limit(96_000)
7            .chunk_size(256)
8            .vision_llm(true)
9            .context_llm(true)
10            .parallel(true)
11            .cache(true),
12    )
13    .add_file("diagram.png")
14    .run()
15    .await?;

Field	Type	Default	Effect
context_limit	usize	128,000	Maximum output token budget.
vision_llm	bool	false	Initializes SmolVLM-256M for image description.
context_llm	bool	false	Initializes SmolLM2-135M for post-selection polishing.
chunk_size	usize	512	Syncs the chunk target token size.
parallel	bool	true	Stored on config; the current builder path still uses parallel file reads regardless.
cache	bool	true	Stored on config; the current global embedding cache is not toggled by this flag yet.

`parallel` and `cache` are part of the public config surface, but in the current source they are stored rather than used to branch runtime behavior.

Local model behavior

1use forgetless::{Config, Forgetless};
2
3// Fast path: embeddings + heuristics only
4let fast = Forgetless::new()
5    .config(Config::default().context_limit(64_000))
6    .add_file("paper.pdf");
7
8// Add local image descriptions
9let with_vision = Forgetless::new()
10    .config(Config::default().vision_llm(true))
11    .add_file("diagram.png");
12
13// Add local text polishing
14let with_context_llm = Forgetless::new()
15    .config(Config::default().context_llm(true))
16    .add(long_context);

Vision mode loads HuggingFaceTB/SmolVLM-256M-Instruct. Context polish mode loads HuggingFaceTB/SmolLM2-135M-Instruct and only runs after chunk selection.

Chunk tuning

1use forgetless::{ChunkConfig, Config, ForgetlessConfig, ScoringConfig};
2
3let advanced = ForgetlessConfig::new(
4    Config::default().context_limit(64_000),
5)
6.with_chunk(
7    ChunkConfig::for_quality()
8        .with_target_tokens(256)
9        .with_max_tokens(512)
10        .with_min_tokens(10),
11)
12.with_scoring(ScoringConfig {
13    semantic_weight: 0.6,
14    keyword_weight: 0.25,
15    priority_weight: 0.15,
16});

Preset	Target	When to use it
ChunkConfig::default()	512	General-purpose prompt compression.
ChunkConfig::for_code()	256	More granular code selection.
ChunkConfig::for_conversation()	200	Message-oriented chat history.
ChunkConfig::for_speed()	1000	Faster processing with larger chunks.
ChunkConfig::for_quality()	256	Smaller chunks for more precise selection.

Tokenizer options

1use forgetless::{Config, ForgetlessConfig, TokenizerModel};
2
3let config = ForgetlessConfig::new(Config::default())
4    .with_tokenizer(TokenizerModel::Custom {
5        chars_per_token_x100: 400,
6    });

The default tokenizer uses cl100k_base. A custom characters-per-token ratio is available when you want a lightweight estimate instead.

Chunking

Tokenizer