better_decoder.py

from __future__ import annotations

import argparse
import csv
import os
import sys
import json
from typing import List, Tuple, Optional, Dict

import torch
import torch.nn.functional as F

from model import HybridTimeScaleConfig
from train_split_model import (
    BinaryBoundaryModel,
    SLP1_CHARS,
    char_to_idx,
    PAD_ID,
    VOCAB_SIZE,
)

from vidyut.sandhi import Splitter
from vidyut.kosha import Kosha, PadaEntry
from vidyut import kosha as kosha_mod, cheda as cheda_mod
from SanskritTokenizer.cache import get_default_data_path
from vidyut.prakriya import Vyakarana
from vidyut import lipi
from sanskrit_parser.base.sanskrit_base import SanskritImmutableString
from sanskrit_parser.parser.sandhi import Sandhi
import logging
logger = logging.getLogger("sanskrit_parser")
logger.setLevel(logging.WARNING)


data_path = get_default_data_path()
rules_csv = os.path.join(data_path, "sandhi", "rules.csv")
_splitter = Splitter.from_csv(rules_csv)
_sandhi = Sandhi()
_kosha = Kosha(data_path+'/kosha')
chedaka = cheda_mod.Chedaka(data_path)
vyakarana = Vyakarana()
import re
_filter = lambda word: [vyakarana.derive(_kosha.get(word)[0])[i].text for i in range(len(vyakarana.derive(_kosha.get(word)[0]))) if vyakarana.derive(_kosha.get(word)[0])[i].text == word]
_make_normal = lambda word: re.sub('s$', '', word)
_filter = lambda word: [vyakarana.derive(_kosha.get(re.sub('M$', 'm', word))[0])[i].text for i in range(len(vyakarana.derive(_kosha.get(re.sub('M$', 'm', word))[0]))) if vyakarana.derive(_kosha.get(re.sub('M$', 'm', word))[0])[i].text == re.sub('M$', 'm', word)]
_check = lambda word: _make_normal(word) == _filter(_make_normal(word))[0]

def predict(model, text: str, delay: int=2) -> torch.Tensor:
    model.eval()
    device = next(model.parameters()).device

    # Build input: real chars + N pad chars at end
    N = delay
    L = len(text)
    total_len = L + N

    input_ids = torch.tensor(
        [[char_to_idx.get(c, PAD_ID) for c in text] + [PAD_ID] * N],
        dtype=torch.long, device=device,
    )
    attention_mask = torch.ones(total_len, dtype=torch.long, device=device).unsqueeze(0)
    # Mark appended pads as valid for model processing (so it can predict
    # boundaries in the last N chars of real text)
    attention_mask[0, L:] = 1  # keep pads in attention

    with torch.no_grad():
        output = model(input_ids=input_ids, attention_mask=attention_mask)
        raw_probs = torch.sigmoid(output["logits"][0]).cpu()  # (total_len,)

    # Shift: output at position bnd+N predicts for position bnd
    # So probs[bnd] = raw_probs[bnd + N]
    probs = torch.zeros(L)
    for bnd in range(L):
        if bnd + N < total_len:
            probs[bnd] = raw_probs[bnd + N]

    return probs

def predicted_pos(model, text, threshold=0.85, delay=2, send_refined=True):
    prob = predict(model, text, delay=delay)
    base_pos = torch.flatten(torch.nonzero(prob > threshold))
    refined_prob = prob_refiner(prob, threshold)
    return refined_prob if send_refined else base_pos, prob

import torch
from typing import List, Union

def prob_refiner(prob: Union[torch.Tensor, List[float]], threshold: float = 0.85) -> torch.Tensor:
    """
    Refines probability tensors into split indices using custom domain heuristics:
    1. Isolated low-probability spikes (e.g. [0, 0, 0.5, 0]) are accepted.
    2. Drops below 50% of the previous value break continuous runs.
    3. Triplets (e.g. [0.9, 0.95, 0.9] or [0.9, 0.7, 0.9]) discard the center and pick outer elements.
    4. Continuous runs >= threshold are grouped in pairs of 2, picking the 2nd index per group.
    
    Args:
        prob: 1D probability sequence (PyTorch Tensor or List).
        threshold: High-confidence threshold (default: 0.85).
        
    Returns:
        1D torch.LongTensor of accepted split candidate indices.
    """
    if isinstance(prob, list):
        prob = torch.tensor(prob, dtype=torch.float32)
    elif not isinstance(prob, torch.Tensor):
        prob = torch.as_tensor(prob, dtype=torch.float32)
        
    n = len(prob)
    if n == 0:
        return torch.tensor([], dtype=torch.long)
        
    candidates = []
    
    # -------------------------------------------------------------
    # Rule 1: Accept isolated spikes significantly larger than 0 surroundings
    # -------------------------------------------------------------
    for i in range(n):
        val = prob[i].item()
        if 0.3 < val < threshold:
            left = prob[i - 1].item() if i > 0 else 0.0
            right = prob[i + 1].item() if i < n - 1 else 0.0
            
            # If both surrounding values are <= half of current value (or zero)
            if left <= (val / 2.0) and right <= (val / 2.0):
                candidates.append(i)

    # -------------------------------------------------------------
    # Rule 2: Collect high-confidence indices and break on >50% relative drops
    # -------------------------------------------------------------
    above_thresh = (prob >= threshold).nonzero(as_tuple=True)[0].tolist()
    
    if not above_thresh:
        return torch.tensor(sorted(list(set(candidates))), dtype=torch.long)

    # Group consecutive indices into continuous blocks
    blocks = []
    current_block = [above_thresh[0]]
    
    for idx in above_thresh[1:]:
        prev_idx = current_block[-1]
        
        # Check if adjacent AND probability hasn't dropped below half of previous
        if idx == prev_idx + 1:
            if prob[idx].item() < (prob[prev_idx].item() * 0.5):
                # Prob dropped below 50% of previous -> end current block & start new block
                blocks.append(current_block)
                current_block = [idx]
            else:
                current_block.append(idx)
        else:
            blocks.append(current_block)
            current_block = [idx]
            
    blocks.append(current_block)

    # -------------------------------------------------------------
    # Rule 3 & 4: Apply Triplet Outer-Selection & Pair Grouping
    # -------------------------------------------------------------
    for block in blocks:
        b_len = len(block)
        
        if b_len == 1:
            candidates.append(block[0])
            
        elif b_len == 2:
            # Pair: accept the later (2nd) candidate
            candidates.append(block[1])
            
        elif b_len == 3:
            # Triplet rule: Pick outer bounds (idx 0 and idx 2), discard middle peak
            candidates.append(block[1])
            
        else:
            # Block length >= 4: Group by 2 for continuous section, pick 2nd per pair
            for i in range(0, b_len, 2):
                pair = block[i:i+2]
                if len(pair) == 2:
                    candidates.append(pair[1])
                else:
                    # Trailing single element for odd-length blocks
                    candidates.append(pair[0])

    # Deduplicate and return sorted indices
    final_candidates = sorted(list(set(candidates)))
    return torch.tensor(final_candidates, dtype=torch.long)


def split_at_indices(text, indices):
    """
    Splits a string into substrings at the specified indices.
    """
    # Sort indices just in case they are out of order
    # Add 0 at the start and None at the end to capture the full string
    split_points = [-1] + sorted(indices) + [len(text)]
    
    # Slice the string between each pair of points
    # The 'if' condition prevents empty strings if an index is out of bounds
    return [
        [text[split_points[i]+1 : min(len(text), split_points[i+1]+1)], split_points[i+1] - split_points[i]]
        for i in range(len(split_points)-1) 
        if text[split_points[i]+1 : min(len(text), split_points[i+1]+1)]
    ]


def give_model_split(model, text, threshold=0.85, delay=2):
    pos, prob = predicted_pos(model, text, threshold, delay)
    return split_at_indices(text, list(pos))


def do_sandhi_join(*arg_words):
    if len(arg_words)==0:
        return ''
    join = arg_words[0]
    for word in arg_words[1:]:
        join_list = sorted(list(_sandhi.join(SanskritImmutableString(join, encoding='slp1'), SanskritImmutableString(word, encoding='slp1'))))
        filtered = [_ords for _ords in join_list if _ords!=(join+word)]
        if len(filtered):
            join = filtered[0]
        else:
            join = join_list[0]
    return join.replace(' ', '')


def verify_sandhi_split(first_word, second_word, compound_word=None):
    if compound_word is None:
        print("compound_word is none")
        return False
    return compound_word in _sandhi.join(SanskritImmutableString(first_word, encoding='slp1'), SanskritImmutableString(second_word, encoding='slp1'))

def give_vidyut_sandhi_split(first_word, second_word, first_index, next_char):
    verified_splits = []
    count = {}
    for delta in [0, -1, 1]:
        if delta == 0:
            compound_word = first_word + second_word
        if delta == -1:
            compound_word = (first_word + second_word)[:-1]
        if delta == 1:
            compound_word = (first_word + second_word) + next_char
        if first_index>=len(compound_word):
            continue
        splits = _splitter.split_at(compound_word, first_index)
        valid_splits = [split for split in splits if split.is_valid]
        for split in valid_splits:
            if verify_sandhi_split(split.first, split.second, compound_word):
                verified_splits.append([split, delta])
    for split,_ in verified_splits:
        count[split.first] = 1 + count.get(split.first, 0)
    first_verified_word = sorted(list(count.items()), key=lambda x: x[-1])[-1][0] if len(verified_splits) else ''
    for split, delta in verified_splits:
        if split.first == first_verified_word:
            return split.first, split.second, delta
    return '', '', None

def give_vidyut_chedaka_split(first_word, second_word, first_index, next_char):
    verified_splits = []
    for delta in [0, -1, 1]:
        if delta == 0:
            compound_word = first_word + second_word
        if delta == -1:
            compound_word = (first_word + second_word)[:-1]
        if delta == 1:
            compound_word = (first_word + second_word) + next_char
        chedaka_run = chedaka.run(compound_word)
        all_words = [_cheda.text for _cheda in chedaka_run]
        sandhi_joined = do_sandhi_join(*all_words) if len(all_words) else ''
        if len(chedaka_run)>1 and sandhi_joined == compound_word:
            verified_splits.append([chedaka_run, delta])
    min_chedaka_count = min([len(cheda_run) for cheda_run, _ in verified_splits]) if len(verified_splits) else 0
    for chedaka_run, delta in verified_splits:
        if len(chedaka_run)==min_chedaka_count:
            return chedaka_run, delta
    return None, None

def do_delta_transform(delta, splited_text, second_word, chedaka_residue=None):
    if delta == 0:
        pass
    elif delta == -1:
        if len(splited_text):
            splited_text[-1][0] = second_word[-1] + splited_text[-1][0]
            splited_text[-1][-1] = splited_text[-1][-1] + 1
    elif delta == 1:
        if len(splited_text):
            splited_text[-1][0] = splited_text[-1][0][1:]
            splited_text[-1][-1] = splited_text[-1][-1] - 1
    if chedaka_residue:
        splited_text[-1][0] = chedaka_residue + splited_text[-1][0]
        splited_text[-1][-1] = splited_text[-1][-1] + len(chedaka_residue)
    return splited_text

def score_word_with_kosha(word: str, kosha: Kosha = _kosha) -> float:
    """
    Scores a candidate Sanskrit string (in SLP1) using Vidyut's dictionary.
    
    Returns:
        1.0 -> Fully inflected word / valid Pada (e.g., 'kamalA', 'uktvA', 'aDikAraH')
        0.5 -> Compound stem / Pratipadika (e.g., 'kamala', 'hita')
        0.0 -> Non-word / Sandhi hallucination (e.g., 'uktva', 'akamala')
    """
    entries = kosha.get(word)
    if word == 'eva':
        print(word)
    # 1. Zero points for non-words
    if not entries:
        return 0.0
    
    # 2. Maximum points for a complete, inflected word (Pada)
    if any(isinstance(entry, PadaEntry) for entry in entries):
        return 1.0
    
    # 3. Partial points for uninflected compound stems
    return 0.5

def score_sandhi_split_candidates(list_candidates, compound_word, external_comp):
    score = {}
    for idx, (first_word_w, second_word_w, _delta) in enumerate(list_candidates):
        first_word = first_word_w[0]
        expected_length = first_word_w[1]
        if first_word not in compound_word:
            score[idx] = score.get(idx, 0) + 0.2
        score[idx] = score.get(idx, 0) - abs((expected_length-len(first_word))/expected_length)
        score[idx] = score.get(idx, 0) - _delta/2
        if first_word == external_comp:
            score[idx] = score.get(idx, 0) + 0.5
        score[idx] = score.get(idx, 0) + score_word_with_kosha(first_word)
    wining_idx = sorted(list(score.items()), key = lambda x: x[1], reverse=True)[0][0] if score else None
    return list_candidates[wining_idx] if wining_idx else [['', None], ['', None], None]

def give_refinement(splited_text):
    refined_result = []
    splited_text.reverse() # for easier poping
    # doing preemptive check to remove model's correct preds
    # preemptive_check = copy.deepcopy(splited_text)[::-1]
    # for _split in preemptive_check:
    #     if _kosha.get(_split[0]):
    #         refined_result.append(splited_text.pop())
    #     else:
    #         break
    # if refined_result and splited_text:
    #     # adding the len zero chck of while loop below to keep consistence with preemptive check
    #     refined_result.append(splited_text.pop())
    while len(splited_text):
        if len(refined_result)==0:
            refined_result.append(splited_text.pop())
            continue

        first_word, first_index = refined_result.pop() # refining last appended word
        second_word, second_index = splited_text.pop()
        first_word = first_word
        second_word = second_word
        if len(splited_text)==0:
            # last word so we will do check complete word completion
            try:
                compound_word = first_word + second_word
                if _make_normal(compound_word) == _make_normal(vyakarana.derive(_kosha.get(compound_word)[0])[0].text):
                    refined_result.append([compound_word, -1])
                    continue
            except:
                pass
        
        next_char = ''
        if len(splited_text):
            next_char = splited_text[-1][0][0]
        chedaka_run, chedaka_delta = give_vidyut_chedaka_split(first_word, second_word, first_index, next_char)
        sandhi_candidates = []
        for index_delta in [-1, -2, 0, -3, 1, 2]:
            if (first_index+index_delta)<0 or (first_index+index_delta)>len(first_word+second_word):
                continue
            sandhi_first, sandhi_second, sandhi_delta = give_vidyut_sandhi_split(first_word, second_word, first_index+index_delta, next_char)

            delta = None
            if chedaka_run is not None:
                chedaka_split_index = len(chedaka_run[0].text)
            else:
                chedaka_split_index = 10000
            if sandhi_delta is not None:
                sandhi_split_index = len(sandhi_first)
            else:
                sandhi_split_index = 10000
            if (sandhi_delta is not None):
                sandhi_candidates.append([[sandhi_first, first_index], [sandhi_second, second_index], sandhi_delta])
        chedana_first = chedaka_run[0].text if chedaka_run else ''
        best_sandhi_candidate = score_sandhi_split_candidates(sandhi_candidates, (first_word+second_word), chedana_first)
        sandhi_delta = best_sandhi_candidate[-1]
        chedaka_residue = None
        if (chedaka_run and do_sandhi_join(first_word, second_word)==do_sandhi_join(*[cheda.text for cheda in chedaka_run])) and \
            (((sandhi_delta is not None) and chedaka_run[0].text == sandhi_first) or (abs(sandhi_split_index-first_index)>abs(chedaka_split_index-first_index))):
            # prefer chedaka over sandhi if same first word
            refined_result.append([chedaka_run[0].text, first_index])
            refined_result.append([chedaka_run[1].text, second_index])
            chedaka_residue = do_sandhi_join(*[cheda.text for cheda in chedaka_run[2:]]) if len(chedaka_run)>2 else None
            delta = chedaka_delta
        elif (sandhi_delta is not None):
            refined_result.append([best_sandhi_candidate[0][0], best_sandhi_candidate[0][1]])
            refined_result.append([best_sandhi_candidate[1][0], best_sandhi_candidate[1][1]])
            delta = sandhi_delta
        if delta is not None:
            splited_text = do_delta_transform(delta, splited_text, second_word, chedaka_residue)
        elif delta is None:
            # we have failed to have any proper split for this compound
            # we will check against shabda kosha if any of the word is available to us
            try:
                second_word_valid = _check(second_word)
            except:
                second_word_valid = False
            if _kosha.get(first_word):
                refined_result.append([first_word, first_index])
                refined_result.append([second_word, second_index])
                continue
            # otherwise we might consider, relabling of the remaining string:
            left_overstring = ''.join([w_i[0] for w_i in splited_text[::-1]])
            remaining_string = first_word + second_word + left_overstring
            # we don;t have access to model right now, so we will send the string to caller and loop the iterations
            refined_result.append([remaining_string, -10])
            return refined_result # return early for handling this case
    return refined_result

def refinement_looper(model, text, threshold=0.85, threshold_limit=0.4):
    _input_text = text
    result = []
    loop_for_now = True
    splits = give_model_split(model, _input_text, threshold)
    while loop_for_now and threshold>threshold_limit: # looping condition
        splits = give_model_split(model, _input_text, threshold)
        splits = give_refinement(splits)
        if len(splits)==1:
            # resolution fail
            # will retry with lower threshold
            threshold -= 0.1
            continue

        loop_for_now = False
        for split in splits:
            if split[-1]!=-10:
                result.append(split[0])
            else:
                # there is left over string
                _input_text = splits[-1][0]
                loop_for_now = True
    return result + ([split[0] for split in splits] if loop_for_now else [])


def refinement_seq(model, text, threshold=0.85, threshold_limit=0.4):
    _inpt = ['asti', text]
    result = []
    while threshold>threshold_limit:
        remaining = do_sandhi_join(*_inpt[1:])
        if len(_inpt)<=2:
            threshold-=0.12
        _inpt=refinement_looper(model, do_sandhi_join(*_inpt[:]), threshold, threshold_limit)
        result.append(_inpt.pop(1)) if (len(_inpt)>1 and _inpt[0]=='asti') else ''
        if _inpt[0] != 'asti' and do_sandhi_join(*_inpt)!='asti':
            _inpt = ['', remaining]
    return result

def give_chedaka_refinement(splited_text):
    refined_result = []
    splited_text.reverse() # for easier poping
    while len(splited_text):
        if len(refined_result)==0:
            refined_result.append(splited_text.pop())
            continue
        # try:
        #     if len(refined_result)>1 and _check(refined_result[-1][0]):
        #         refined_result.append(splited_text.pop())
        # except:
        #     pass
        first_word, first_index = refined_result.pop() # refining last appended word
        second_word, second_index = splited_text.pop()
        compound_word = first_word+second_word
        if len(splited_text)==0:
            # last word so we will do check complete word completion
            try:
                if _make_normal(compound_word) == _make_normal(vyakarana.derive(_kosha.get(compound_word)[0])[0].text):
                    refined_result.append([compound_word, -1])
                    continue
            except:
                pass
        for delta in [0, -1, 1]:
            if delta == 0:
                chedaka_run = chedaka.run(compound_word)
            elif delta == -1:
                chedaka_run = chedaka.run(compound_word[:delta])
            elif delta == 1:
                if len(splited_text):
                    next_char = splited_text[-1][0][0]
                else:
                    next_char = ''
                chedaka_run = chedaka.run(compound_word+next_char)
            if len(chedaka_run)>1:
                chedaka_split_idx = len(chedaka_run[0].text)
                sandhi_splits = _splitter.split_at(compound_word, min(len(compound_word)-1, first_index-1))
                valid_sandhi_splits = [split for split in sandhi_splits if split.is_valid]
                if len(valid_sandhi_splits):
                    list_candidates = [[[split.first, first_index], [split.second, second_index], -1] for split in valid_sandhi_splits]
                    ((sandhi_first, _), (sandhi_second, _), is_score_not_none) = score_sandhi_split_candidates(list_candidates, compound_word, chedaka_run[0].text)
                    sandhi_split = [sandi for sandi in valid_sandhi_splits if (sandi.first==sandhi_first and sandi.second==sandhi_second)][0] if is_score_not_none  else None
                    sandhi_split_index = len(sandhi_split.first) if sandhi_split else None
                else:
                    sandhi_split_index = None
                    sandhi_split = None
                if sandhi_split_index is None or abs(sandhi_split_index-first_index)>abs(chedaka_split_idx-first_index):
                    for _cheda in chedaka_run:
                        refined_result.append([_cheda.text, second_index]) # this losses first index info but okay for simplicity, only last index matters
                else:
                    try:
                        if _check(sandhi_split.first):
                            refined_result.append([sandhi_split.first, first_index])
                            refined_result.append([sandhi_split.second, second_index])
                        elif _check(chedaka_run[0].text):
                            for _cheda in chedaka_run:
                                refined_result.append([_cheda.text, second_index]) # this losses first index info but okay for simplicity, only last index matters
                    except:
                        try:
                            if _check(chedaka_run[0].text):
                                for _cheda in chedaka_run:
                                    refined_result.append([_cheda.text, second_index]) # this losses first index info but okay for simplicity, only last index matters
                        except:
                            refined_result.append([sandhi_split.first, first_index])
                            refined_result.append([sandhi_split.second, second_index])
                if delta == 0:
                    pass
                elif delta == -1:
                    if len(splited_text):
                        splited_text[-1][0] = compound_word[-1] + splited_text[-1][0]
                elif delta == 1:
                    if len(splited_text):
                        splited_text[-1][0] = splited_text[-1][0][1:]
                break
        if len(chedaka_run)<=1:
            for delta in [0, -1, 1]:
                if delta == 0:
                    sandhi_splits = _splitter.split_at(compound_word, first_index-1)
                    valid_sandhi_splits = [split for split in sandhi_splits if split.is_valid]
                    sandhi_split = valid_sandhi_splits[0] if len(valid_sandhi_splits) else None
                    sandhi_split_index = len(sandhi_split.first) if sandhi_split else None
                elif delta == -1:
                    sandhi_splits = _splitter.split_at(compound_word[:delta], first_index-1)
                    valid_sandhi_splits = [split for split in sandhi_splits if split.is_valid]
                    sandhi_split = valid_sandhi_splits[0] if len(valid_sandhi_splits) else None
                    sandhi_split_index = len(sandhi_split.first) if sandhi_split else None
                elif delta == 1:
                    if len(splited_text):
                        next_char = splited_text[-1][0][0]
                    else:
                        next_char = ''
                    sandhi_splits = _splitter.split_at(compound_word+next_char, first_index-1)
                    valid_sandhi_splits = [split for split in sandhi_splits if split.is_valid]
                    sandhi_split = valid_sandhi_splits[0] if len(valid_sandhi_splits) else None
                    sandhi_split_index = len(sandhi_split.first) if sandhi_split else None
                if sandhi_split_index is not None:
                    if delta == 0:
                        pass
                    elif delta == -1:
                        if len(splited_text):
                            splited_text[-1][0] = second_word[-1] + splited_text[-1][0]
                    elif delta == 1:
                        if len(splited_text):
                            splited_text[-1][0] = splited_text[-1][0][1:]
                    break
            if sandhi_split_index is not None:
                refined_result.append([sandhi_split.first, first_index])
                refined_result.append([sandhi_split.second, second_index])
            else:
                refined_result.append([compound_word, second_index])

    return refined_result

def give_split(text, id, skip=0):
    _segmentation = _splitter.split_at(text[skip:], id - skip)[0]
    if _segmentation.is_valid:
        return _segmentation.first

def get_best_vidyut_split(text, split_idx, neural_prob):
    candidates = _splitter.split_at(text, split_idx)
    
    best_candidate = None
    best_score = -1.0
    
    for cand in candidates:
        # 1. Grammar is King: Does Vidyut recognize the resulting words in its lexicon?
        is_lexically_valid = cand.is_valid
        
        # Heavy penalty for proposing non-words (e.g., 'uktva' gets crushed here)
        base_score = 1.0 if is_lexically_valid else 0.1
        
        # 2. Smart Sandhi Evaluation
        # Check if the split required changing letters (Sandhi inversion)
        is_sandhi_inversion = (cand.first != text[:len(cand.first)])
        
        if is_sandhi_inversion:
            if is_lexically_valid:
                # REWARD: It's a real dictionary word that required un-joining (e.g., hito -> hita)
                base_score += 0.2
            else:
                # PENALIZE: It's hallucinating fake sandhi rules to force a split (e.g., uktvA -> uktva)
                base_score -= 0.05
        else:
            if is_lexically_valid:
                # SLIGHT REWARD: Valid literal cuts (like uktvA + kamalA) are highly reliable
                base_score += 0.1
        
        if is_lexically_valid:
               base_score += score_word_with_kosha(cand.second)

        # 3. Combine with Neural Confidence
        total_score = base_score * neural_prob
        
        if total_score > best_score:
            best_score = total_score
            best_candidate = cand
            
    return best_candidate


def get_score_cand_nxt(text, cand, next_word):
    base_score = 0
    # 1. Grammar is King: Does Vidyut recognize the resulting words in its lexicon?
    is_lexically_valid = cand.is_valid
    
    # Heavy penalty for proposing non-words (e.g., 'uktva' gets crushed here)
    base_score = 1.0 if is_lexically_valid else 0.1
    
    # 2. Smart Sandhi Evaluation
    # Check if the split required changing letters (Sandhi inversion)
    is_sandhi_inversion = (cand.first != text[:len(cand.first)])
    
    if is_sandhi_inversion:
        if is_lexically_valid:
            # REWARD: It's a real dictionary word that required un-joining (e.g., hito -> hita)
            base_score += 0.3
        else:
            # PENALIZE: It's hallucinating fake sandhi rules to force a split (e.g., uktvA -> uktva)
            base_score -= 0.05
    else:
        if is_lexically_valid:
            # SLIGHT REWARD: Valid literal cuts (like uktvA + kamalA) are highly reliable
            base_score += 0.1
    
    if is_lexically_valid:
        base_score += score_word_with_kosha(cand.second)
        base_score += score_word_with_kosha(next_word)

    return base_score


def get_best_with_next_word(model, text, original_split_idx, neural_prob):
   candidate = get_best_vidyut_split(text, original_split_idx, neural_prob)
   next_best_guess_word = sequential_backward_split(model, candidate.second)[0] if candidate.is_valid and candidate.second else ''
   return candidate, next_best_guess_word

def get_searched_vidyut_split(text, original_split_idx, neural_prob, minus_1_prob, model=None):
    if model is None:
        plus_1_cand = get_best_vidyut_split(text, original_split_idx+1, neural_prob)
        orginal_cand = get_best_vidyut_split(text, original_split_idx, neural_prob)
        minus_1_cand = get_best_vidyut_split(text, original_split_idx-1, neural_prob) if original_split_idx-1>0 else orginal_cand
        if minus_1_cand.is_valid and minus_1_prob>0.3:
            return minus_1_cand
        if neural_prob<0.95 and plus_1_cand.is_valid:
            return plus_1_cand
        return orginal_cand
    else:
        score_orginal_cand = 0
        score_plus_cand = 0
        score_minus_cand = 0
        orginal_cand, next_org_word = get_best_with_next_word(model, text, original_split_idx, neural_prob)
        plus_1_cand, next_plus_word = get_best_with_next_word(model, text, original_split_idx+1, neural_prob) if original_split_idx+1<len(text) else (orginal_cand, next_org_word)
        minus_1_cand, next_minus_word = get_best_with_next_word(model, text, original_split_idx-1, neural_prob) if original_split_idx-1>0 else (orginal_cand, next_org_word)
        plus_score = get_score_cand_nxt(text, plus_1_cand, next_plus_word)
        org_score = get_score_cand_nxt(text, orginal_cand, next_org_word)
        minus_score = get_score_cand_nxt(text, minus_1_cand, next_minus_word)
        score_word = [[plus_score, plus_1_cand], [org_score, orginal_cand], [minus_score, minus_1_cand]]
        return sorted(score_word, reverse=True)[0][-1]

def sequential_split(model, text, threshold=0.85, delay=2):
    result_texts = []
    while text!="": 
        split_candidate = give_refinement(give_model_split(model, text))[0][0]
        print(split_candidate)
        result_texts.append(split_candidate)
        valid_ = [v for v in  _splitter.split_at(text, max(0, len(split_candidate)-1)) if v.is_valid]
        text = valid_[0].second if valid_ else _splitter.split_at(text, max(0, len(split_candidate)-1))[0].second
    return result_texts

def sequential_backward_split(model, text, threshold=0.85, delay=2):
    result_texts = []
    positions_split, prob_split = predicted_pos(model, text, threshold=threshold, delay=delay) # init values
    while text!="" and len(positions_split)>0:
        positions_split, prob_split = predicted_pos(model, text, threshold=threshold, delay=delay)
        positions_split = list(positions_split)
        while len(positions_split):
            pos_split = positions_split.pop()
            split_candidate = get_searched_vidyut_split(text, pos_split, prob_split[pos_split], prob_split[max(0, pos_split-1)], model=model)
            if split_candidate.is_valid and split_candidate.second:
                result_texts.append(split_candidate.second)
                # print("splited : ", split_candidate.first)
                # print(len(positions_split))
                text = split_candidate.first
                break # for chunking remainging compound
            else:
                pass # skip unvieriied splits for now simple
        # print(text, positions_split)
    if text!='':
        result_texts.append(text)
    result_texts.reverse()
    return result_texts

# ---------------------------------------------------------------------------
# Load model
# ---------------------------------------------------------------------------

def load_model(checkpoint_dir: str, device: torch.device) -> Tuple[BinaryBoundaryModel, int]:
    """Load model and extract delay from config or checkpoint."""
    config_path = os.path.join(checkpoint_dir, "config.json")
    if os.path.exists(config_path):
        config = HybridTimeScaleConfig.from_pretrained(checkpoint_dir)
    else:
        # Infer config from checkpoint weights
        print(f"[warn] No config.json, inferring from checkpoint", file=sys.stderr)
        from safetensors.torch import load_file
        sd_path = os.path.join(checkpoint_dir, "model.safetensors")
        if not os.path.exists(sd_path):
            sd_path = os.path.join(checkpoint_dir, "pytorch_model.bin")
            sd = torch.load(sd_path, map_location="cpu", weights_only=False)
        else:
            sd = load_file(sd_path)

        # Clean keys
        sd = {k.replace("_orig_mod.", ""): v for k, v in sd.items()}

        latent_dim = sd["embedding.weight"].shape[1]
        q_out = sd["blocks.0.mixer.q_proj.weight"].shape[0]
        # num_heads defaults to max(1, latent_dim // 64) in HybridSpectralBlock
        num_heads = max(1, latent_dim // 64)
        num_modes = q_out // num_heads
        num_layers = sum(1 for k in sd if k.startswith("blocks.") and k.endswith(".ffn.1.weight"))

        print(f"  Inferred: latent_dim={latent_dim}, num_modes={num_modes}, "
              f"num_heads={num_heads}, num_layers={num_layers}", file=sys.stderr)

        config = HybridTimeScaleConfig(
            vocab_size=VOCAB_SIZE, latent_dim=latent_dim, num_layers=num_layers,
            num_modes=num_modes,
            layer_types=["linear", "softmax"][:num_layers] or ["linear", "softmax"],
            time_scale=float(2 * num_modes),
            dropout=0.1, pad_token_id=PAD_ID,
            bos_token_id=0, eos_token_id=0, tie_word_embeddings=False,
        )

    delay = getattr(config, "delay", 0)
    print(f"Model delay: {delay}", file=sys.stderr)

    model = BinaryBoundaryModel(config)
    sd_path = os.path.join(checkpoint_dir, "model.safetensors")
    if os.path.exists(sd_path):
        from safetensors.torch import load_file
        state_dict = load_file(sd_path)
    else:
        sd_path = os.path.join(checkpoint_dir, "pytorch_model.bin")
        state_dict = torch.load(sd_path, map_location="cpu", weights_only=False)
    state_dict = {k.replace("_orig_mod.", ""): v for k, v in state_dict.items()}
    model.load_state_dict(state_dict, strict=False)
    model = model.to(device)
    model.eval()
    return model, delay


import copy
def better_decode(model, text):
    _input = text
    result = []
    splits = give_refinement(give_model_split(model, copy.deepcopy(_input), 0.3))
    while _input or len(splits)>1:
        _input = text[len(splits[0][0]):]
        result.append(splits[0][0])
        splits = give_refinement(give_model_split(model, copy.deepcopy(_input), 0.3))
    return result + [li[0] for li in splits]


def give_correct_slp1_input(text):
    return do_sandhi_join(*lipi.transliterate(text, lipi.Scheme.Devanagari, lipi.Scheme.Slp1).split())

def main():
    parser = argparse.ArgumentParser(description="Run Sandhi vigraha test suite.")
    parser.add_argument("text", nargs="?", default=None)
    parser.add_argument("--batch", default=None)
    parser.add_argument("--test-file", default="test_cases.json")
    parser.add_argument("--checkpoint", default="checkpoints/delayed_model_save_fixed_delay2_v2")
    parser.add_argument("--delay", type=int, default=2)
    parser.add_argument("--threshold", type=float, default=0.85)
    parser.add_argument("--bias-first-word", action="store_true")
    parser.add_argument("--verbose", "-v", action="store_true")
    args = parser.parse_args()

    if torch.cuda.is_available():
        device = torch.device("cuda")
    elif torch.backends.mps.is_available():
        device = torch.device("mps")
    else:
        device = torch.device("cpu")
    print(f"Device: {device}", file=sys.stderr)

    # Load model
    print(f"Loading model from {args.checkpoint}...", file=sys.stderr)
    model, delay = load_model(args.checkpoint, device)
    if args.delay is not None:
        delay = args.delay
    print(f"Model loaded (delay={delay})", file=sys.stderr)
    threshold = args.threshold
    # Load test cases
    with open(args.test_file) as f:
        test_cases = json.load(f)
    print(f"\nLoaded {len(test_cases)} test cases\n", file=sys.stderr)
    text = 'vidyAdadAtivinayaMvinayAdyAtipAtratAm'
    res = refinement_seq(model, text)
    print(res)
    give_chedaka_refinement(give_model_split(model, 'karmARyevADikArastemAphalezukadAcana', 0.7))
    give_refinement(give_model_split(model, 'karmARyevADikArastemAphalezukadAcana', 0.4))
    if test_cases:
        for test_case in test_cases:
            text = test_case.get('input', '')
            expected_output = test_case.get('expected', '').split()
            pred_output = sequential_split(model, text, threshold=threshold)
            print("Input Text:\t", text)
            print('Expected Output:\t', expected_output)
            print('Predicted Output:\t', pred_output)
            print()
    
    texts: List[str] = []
    if args.batch:
        with open(args.batch) as f:
            texts = [line.strip() for line in f if line.strip()]
    elif args.text:
        texts = [args.text]
    else:
        print("\nEnter Sanskrit text (SLP1), Ctrl+D to exit:", file=sys.stderr)
        while True:
            try:
                t = input(">>> ").strip()
                if t:
                    print(sequential_split(model, t, threshold=threshold))
                    texts.append(t)
            except EOFError:
                break
    for text in texts:
        print(sequential_split(model, text, threshold=threshold))


if __name__ =='__main__':
    main()