entelix-tokenizer-tiktoken 0.5.5

entelix vendor-accurate TokenCounter for OpenAI BPE encodings (cl100k_base, o200k_base, p50k_base, r50k_base) — wraps tiktoken-rs with eager preload
Documentation

entelix-tokenizer-tiktoken

Vendor-accurate TokenCounter for OpenAI's BPE tokenizer family — cl100k_base, o200k_base, p50k_base, r50k_base. Wraps tiktoken-rs with eager BPE preload at construction so the per-call count stays synchronous per the TokenCounter contract.

[dependencies]
entelix-core = "0.4"
entelix-tokenizer-tiktoken = "0.4"

use std::sync::Arc;
use entelix_core::{ChatModelConfig, TokenCounter};
use entelix_tokenizer_tiktoken::{TiktokenCounter, TiktokenEncoding};

let counter: Arc<dyn TokenCounter> =
    Arc::new(TiktokenCounter::for_encoding(TiktokenEncoding::O200kBase)?);

let config = ChatModelConfig::default().with_token_counter(Arc::clone(&counter));

assert_eq!(counter.count("Hello world"), 2);

Encoding to model mapping

Encoding Models
Cl100kBase GPT-3.5-turbo, GPT-4, GPT-4-turbo, text-embedding-3-*
O200kBase GPT-4o, GPT-4o-mini, o1, o3, o3-mini, o4
P50kBase GPT-3 davinci, codex
R50kBase GPT-3 ada / babbage / curie, GPT-2

The mapping is left to operators by design — OpenAI changes it over time, and accidentally pinning to a stale mapping silently miscounts without surfacing a build failure. Pick the encoding for your target model from the table above and the wrapper preloads the matching BPE tables.

Position in the workspace

Companion crate to entelix-core — TokenCounter is the trait surface; this crate provides one of several concrete impls. Sister crates ship the HuggingFace Tokenizers backend (entelix-tokenizer-hf) and locale-aware morphological counters (Korean / Japanese).

The zero-dependency entelix_core::ByteCountTokenCounter ships in core for development scaffolding; production multilingual workloads should swap in this crate (or a sibling) at ChatModelConfig::with_token_counter.

License

MIT