/**
 * kosha-discovery — Tokenizer-family inference for API-served models.
 *
 * Local runtimes (Ollama, llama.cpp) expose the tokenizer family via
 * {@link LocalRuntimeMetadata.tokenizerFamily}. For API-served models the
 * provider rarely publishes it explicitly, so we derive it from the
 * origin provider plus the model ID. The result is a best-effort hint
 * intended for tokenizer-aware downstream consumers (compression
 * libraries, routing policies, cost estimators) — not a guarantee.
 * @module
 */
/**
 * Infer a best-effort tokenizer-family identifier for a model.
 *
 * Returns a normalized string when the mapping is confident, or
 * `undefined` when no safe inference is available. Consumers must
 * treat the return value as a hint and fall back to their own
 * defaults when it is absent.
 *
 * Known families returned:
 * - `"o200k_base"` — OpenAI GPT-4o, GPT-4.1, o1/o3/o4 families
 * - `"cl100k_base"` — OpenAI GPT-4, GPT-3.5-turbo, text-embedding-3
 * - `"claude"` — Anthropic Claude family (proprietary tokenizer)
 * - `"gemini"` — Google Gemini family (proprietary tokenizer)
 * - `"llama4"` — Meta Llama 4 family (~200k vocab BPE)
 * - `"llama3"` — Meta Llama 3.x family (128k vocab BPE)
 * - `"llama2"` — Meta Llama 2 / CodeLlama (32k vocab)
 * - `"mistral"` — Mistral/Mixtral family
 * - `"cohere"` — Cohere Command/Embed family
 * - `"deepseek"` — DeepSeek family
 * - `"qwen"` — Alibaba Qwen family
 *
 * @param originProvider - The model's creator (not the serving layer).
 *                         Examples: `"anthropic"`, `"openai"`, `"meta"`.
 * @param modelId        - Provider-canonical model ID.
 * @returns Tokenizer-family hint, or `undefined`.
 */
export declare function inferTokenizerFamily(originProvider: string | undefined, modelId: string): string | undefined;
//# sourceMappingURL=tokenizer-family.d.ts.map