LightMem/src/lightmem/factory/topic_segmenter/identity.py

from __future__ import annotations

from typing import Any, Dict, List, Optional

import tiktoken


class _FallbackTokenizer:
    def encode(self, text: str):
        return str(text).split()


class IdentitySegmenter:
    """
    A no-op segmenter that preserves one-input-message-per-segment behavior.
    This keeps LightMem's add-memory pipeline intact while avoiding routine
    topic splitting for benchmarks that require one text turn semantics.
    """

    def __init__(self, config: Optional[Dict[str, Any]] = None, shared: bool = False, compressor=None):
        del shared, compressor
        self.config = config or {}
        self.buffer_len = int(self.config.get("buffer_len", 200000))
        tokenizer_name = self.config.get("tokenizer_name", "o200k_base")
        try:
            self.tokenizer = tiktoken.encoding_for_model(tokenizer_name)
        except Exception:
            try:
                self.tokenizer = tiktoken.get_encoding("o200k_base")
            except Exception:
                self.tokenizer = _FallbackTokenizer()

    def propose_cut(self, buffer_texts: List[str]) -> List[int]:
        del buffer_texts
        return []