code/spec_rag/config.py

from __future__ import annotations

from dataclasses import dataclass


@dataclass(frozen=True)
class IndexConfig:
    dim: int
    m: int = 32
    ef_construction: int = 200
    ef_search: int = 128
    metric: str = "cosine"  # "cosine" or "l2"


@dataclass(frozen=True)
class EmbedConfig:
    model_name: str
    pooling: str = "cls"  # "cls" or "mean"
    max_length: int = 256
    batch_size: int = 64
    normalize: bool = True


@dataclass(frozen=True)
class ProjectorConfig:
    input_dim: int = 768
    llm_dim: int = 4096


# --- Spectra-Reason-GCD (implementation2.md) ---


@dataclass(frozen=True)
class PerceiverResamplerConfig:
    d_model: int = 1024
    llm_dim: int = 4096
    num_latents: int = 64
    num_heads: int = 8
    num_layers: int = 2
    dropout: float = 0.1


@dataclass(frozen=True)
class SpectraReasonGCDConfig:
    """Config for Spectra-Reason-GCD pipeline."""
    llm_name: str = "meta-llama/Meta-Llama-3-8B-Instruct"
    llm_dim: int = 4096
    num_soft_tokens: int = 64
    max_peaks: int = 60
    perceiver: PerceiverResamplerConfig = PerceiverResamplerConfig()