harn-stdlib 0.10.122

// @harn-entrypoint-category llm.stdlib
/**
 * One token in an exact model vocabulary. The tokenizer identity is part of
 * the value so an integer from one vocabulary cannot silently reach another.
 * `text` is nil when one token is only a fragment of a UTF-8 character.
 */
pub type TokenRef = {
  _type: "llm_token",
  id: int,
  tokenizer: string,
  bytes: list<int>,
  text: string?,
}

/** A bounded bias attached to one exact token. */
pub type TokenBias = {token: TokenRef, bias: float}

/**
 * Tokenize text with the exact vocabulary Harn owns for `model`.
 * Approximate token counters cannot create TokenRef values.
 *
 * @effects: []
 * @errors: ["The model has no exact local tokenizer"]
 */
pub fn tokenize(text: string, model: string) -> list<TokenRef> {
  const raw_tokens: list<dict> = tiktoken_encode_tokens(text, model)
  let tokens: list<TokenRef> = []
  for raw in raw_tokens {
    const id = to_int(raw?.id)
    const tokenizer = raw?.tokenizer
    if raw?._type != "llm_token" || id == nil || id < 0 || type_of(tokenizer) != "string" {
      throw "tokenize: host tokenizer returned an invalid token reference"
    }
    let bytes: list<int> = []
    for value in raw?.bytes ?? [] {
      const byte = to_int(value)
      if byte == nil || byte < 0 || byte > 255 {
        throw "tokenize: host tokenizer returned an invalid token byte"
      }
      bytes = bytes + [byte]
    }
    let token_text: string? = nil
    if raw?.text != nil {
      if type_of(raw?.text) != "string" {
        throw "tokenize: host tokenizer returned invalid token text"
      }
      token_text = to_string(raw?.text)
    }
    const token: TokenRef = {
      _type: "llm_token",
      id: id,
      tokenizer: to_string(tokenizer),
      bytes: bytes,
      text: token_text,
    }
    tokens = tokens + [token]
  }
  return tokens
}

/**
 * Decode an ordered token sequence. Every token must carry the same exact
 * tokenizer identity.
 *
 * @effects: []
 * @errors: ["The sequence mixes vocabularies or is not valid UTF-8"]
 */
pub fn detokenize(tokens: list<TokenRef>) -> string {
  return tiktoken_decode_tokens(tokens)
}

/**
 * Construct a token reference when an integration already has an exact token
 * ID. This operation cannot prove that the ID belongs to the named tokenizer;
 * prefer `tokenize` whenever text is available. The tokenizer identity is
 * still mandatory and is checked against the final LLM route.
 *
 * @effects: []
 * @errors: ["id must be non-negative", "tokenizer must name an exact vocabulary"]
 */
pub fn unsafe_token_ref(id: int, tokenizer: string) -> TokenRef {
  if id < 0 {
    throw "unsafe_token_ref: id must be non-negative"
  }
  if !starts_with(tokenizer, "tiktoken:") {
    throw "unsafe_token_ref: tokenizer must be an exact `tiktoken:<encoder>` identity"
  }
  return {_type: "llm_token", id: id, tokenizer: tokenizer, bytes: [], text: nil}
}