File size: 26,160 Bytes
7c7111f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
"""
Vortex-Embed v2 — Retrieval-optimized LF4 static embedding model.

Built on VTXAI/Vortex-Embed-4.7M (4-bit LF4 weights, 29 KB tokenizer).
All training-free upgrades: SIF IDF weighting, top-K principal component
removal, file-path header injection, and search-time file-extension score
bias.

Key results (Webscout codebase, 5,168 chunks, 51 hand-verified queries):

    R@1  = 0.745  (baseline LF4: 0.314, +137%)
    R@5  = 0.843
    R@10 = 0.882
    MRR  = 0.779

Drop-in replacement for `LF4StaticEmbedding` from the v1 model. Same
weight format, same tokenizer, same embed dimension. New arguments are
optional and default to the v2 best configuration.
"""
from __future__ import annotations

import json
import math
import re
from dataclasses import dataclass, field
from pathlib import Path
from typing import List, Optional, Sequence, Tuple, Union

import numpy as np
from safetensors.numpy import load_file, save_file

try:
    from tokenizers import Tokenizer
except Exception:  # pragma: no cover
    Tokenizer = None  # type: ignore[assignment]


# ---------------------------------------------------------------------------
# Path header helpers
# ---------------------------------------------------------------------------

_PATH_SEP_RE = re.compile(r"[_\-\.]+")


def _path_to_header_tokens(path: str) -> List[str]:
    """Snake/kebab/dot-split a file path into semantic tokens.

    Returns the deduped list of directory parts + stem (with the file
    extension stripped from the last part). Order is preserved.

    Example:
        "llm4free/search/engines/duckduckgo_main.py"
        -> ["llm4free", "search", "engines", "duckduckgo", "main"]
    """
    p = Path(path)
    parts = list(p.parts)
    if parts and parts[0].startswith("."):
        parts = parts[1:]
    stem = p.stem
    parts.append(stem)
    suffix = p.suffix.lstrip(".").lower()
    out: List[str] = []
    for part in parts:
        for w in _PATH_SEP_RE.split(part):
            wl = w.lower()
            if wl and wl != suffix:
                out.append(wl)
    seen, dedup = set(), []
    for w in out:
        if w not in seen:
            seen.add(w)
            dedup.append(w)
    return dedup


# ---------------------------------------------------------------------------
# Main model
# ---------------------------------------------------------------------------


@dataclass
class VortexEmbedConfig:
    """Configuration container mirroring the on-disk ``config.json``."""

    vocab_size: int = 29528
    embedding_dim: int = 256
    block_size: int = 32
    num_blocks: int = 8
    model_type: str = "vortex-embed"
    architectures: List[str] = field(default_factory=lambda: ["VortexEmbedV2"])
    # v2-specific retrieval knobs (also persisted to config.json on save)
    sif_a: float = 1e-4
    sif_pc: float = 1.0
    pc_k: int = 8
    header_repeat: int = 15
    py_bonus: float = 0.05
    md_penalty: float = -0.02
    bias_top_k: int = 50
    quantization: str = "lf4"
    bits: int = 4

    @classmethod
    def from_dict(cls, d: dict) -> "VortexEmbedConfig":
        # Accept arbitrary v1 keys; fall back to defaults for unknown ones
        kw = {k: d[k] for k in d if k in cls.__dataclass_fields__}
        return cls(**kw)

    def to_dict(self) -> dict:
        return {k: getattr(self, k) for k in self.__dataclass_fields__}


class VortexEmbedV2:
    """Vortex-Embed v2 — retrieval-optimized LF4 static embedding.

    Pipeline at encode time (per chunk text):
        1. Augment: prepend path-header tokens × ``header_repeat``
        2. Tokenize (HuggingFace fast tokenizer, same as v1)
        3. SIF IDF weighting on every token
        4. Sum tokens per chunk via ``torch.scatter_add_`` (CPU)
        5. Divide by SIF-weighted count
        6. Remove top-``pc_k`` principal components (fitted on corpus)
        7. L2-normalize

    Pipeline at search time (per query):
        1. Encode query with the same pipeline
        2. Cosine score against the index (``qn @ index.T``)
        3. Within top-``bias_top_k`` candidates, add a small per-extension
           score bias (``+py_bonus`` for .py, ``+md_penalty`` for .md) to
           break the ties where README.md / docs/*.md outrank code

    Args:
        packed: ``uint8`` (vocab, dim//2) packed 4-bit weights.
        scales: ``float16`` (vocab, num_blocks) per-block scales.
        zeros:  ``float16`` (vocab, num_blocks) per-block zero-points.
        tokenizer_data: path to ``tokenizer.json`` or its raw JSON string.
        config: configuration dict (or :class:`VortexEmbedConfig`).
        precompute: if True, dequantize the full table to FP32 at load.
    """

    def __init__(
        self,
        packed: np.ndarray,
        scales: np.ndarray,
        zeros: np.ndarray,
        tokenizer_data: Union[str, Path],
        config: Union[dict, VortexEmbedConfig],
        *,
        precompute: bool = True,
    ) -> None:
        self.packed = np.asarray(packed, dtype=np.uint8)
        self.scales = np.asarray(scales, dtype=np.float16)
        self.zeros = np.asarray(zeros, dtype=np.float16)
        self.tokenizer_data = str(tokenizer_data)
        if isinstance(config, dict):
            self.config = VortexEmbedConfig.from_dict(config)
        else:
            self.config = config
        self.vocab_size = int(self.config.vocab_size)
        self.dim = int(self.config.embedding_dim)
        self.block_size = int(self.config.block_size)
        self.num_blocks = int(self.config.num_blocks)
        # v2 retrieval knobs
        self.sif_a = float(self.config.sif_a)
        self.sif_pc = float(self.config.sif_pc)
        self.pc_k = int(self.config.pc_k)
        self.header_repeat = int(self.config.header_repeat)
        self.py_bonus = float(self.config.py_bonus)
        self.md_penalty = float(self.config.md_penalty)
        self.bias_top_k = int(self.config.bias_top_k)
        # State
        self._tokenizer: Optional[Tokenizer] = None
        self._embedding_table: Optional[np.ndarray] = None
        self._sif_weights: Optional[np.ndarray] = None
        self._pc_directions: Optional[np.ndarray] = None
        self._file_paths: Optional[List[str]] = None
        self._chunk_is_py: Optional[np.ndarray] = None
        self._chunk_is_md: Optional[np.ndarray] = None
        self.cache_path: Optional[Path] = None
        if precompute:
            self._embedding_table = self._dequantize_all()

    # ---- properties -----------------------------------------------------

    @property
    def tokenizer(self) -> Tokenizer:
        if self._tokenizer is None:
            if Tokenizer is None:
                raise RuntimeError("tokenizers is required: install via `pip install tokenizers`")
            self._tokenizer = Tokenizer.from_file(self.tokenizer_data)
        return self._tokenizer

    @property
    def embedding_table(self) -> np.ndarray:
        if self._embedding_table is None:
            self._embedding_table = self._dequantize_all()
        return self._embedding_table

    @property
    def model_size_mb(self) -> float:
        if self._embedding_table is not None:
            return self._embedding_table.nbytes / 1e6
        return (self.packed.nbytes + self.scales.nbytes + self.zeros.nbytes) / 1e6

    # ---- (de)serialization ---------------------------------------------

    @classmethod
    def from_pretrained(
        cls,
        path_or_id: Union[str, Path],
        *,
        precompute: bool = True,
        cache_path: Optional[Union[str, Path]] = None,
        **overrides,
    ) -> "VortexEmbedV2":
        """Load from a local model directory or Hugging Face Hub id.

        Expected files in the directory:
            - ``model.safetensors`` (LF4 packed weights)
            - ``config.json`` (model + retrieval config)
            - ``tokenizer.json``
        """
        path = Path(path_or_id)
        if not path.is_dir():
            from huggingface_hub import snapshot_download
            path = Path(snapshot_download(str(path_or_id)))
        tensors = load_file(str(path / "model.safetensors"))
        config = json.loads((path / "config.json").read_text())
        # Apply overrides (e.g. sif_a=1e-3, header_repeat=10, disable bias...)
        for k, v in overrides.items():
            if k in VortexEmbedConfig.__dataclass_fields__:
                config[k] = v
        obj = cls(
            packed=tensors["embedding_packed"],
            scales=tensors["embedding_scales"],
            zeros=tensors["embedding_zeros"],
            tokenizer_data=str(path / "tokenizer.json"),
            config=config,
            precompute=precompute,
        )
        if cache_path is not None:
            obj.cache_path = Path(cache_path)
        return obj

    def save_pretrained(self, path: Union[str, Path]) -> None:
        """Save weights + config + tokenizer to a local directory."""
        out = Path(path)
        out.mkdir(parents=True, exist_ok=True)
        save_file(
            {
                "embedding_packed": self.packed,
                "embedding_scales": self.scales,
                "embedding_zeros": self.zeros,
            },
            str(out / "model.safetensors"),
        )
        (out / "config.json").write_text(
            json.dumps(self.config.to_dict(), indent=2)
        )
        if not (out / "tokenizer.json").exists():
            (out / "tokenizer.json").write_text(
                Path(self.tokenizer_data).read_text()
            )

    # ---- LF4 dequantization --------------------------------------------

    def _dequantize_all(self) -> np.ndarray:
        """Dequantize the complete LF4 embedding table to FP32.

        Each output row is a 256-dim vector. Block-wise: for block b,
        value = scale[b] * int4 + zero[b]. Int4 is stored as 2 nibbles
        per byte (low / high).
        """
        low = (self.packed & 0x0F).astype(np.float32)
        high = ((self.packed >> 4) & 0x0F).astype(np.float32)
        padded = self.packed.shape[1] * 2
        unpacked = np.empty((self.packed.shape[0], padded), dtype=np.float32)
        unpacked[:, 0::2] = low
        unpacked[:, 1::2] = high
        blocked = unpacked.reshape(self.packed.shape[0], self.num_blocks, self.block_size)
        scales = self.scales.astype(np.float32)[:, :, None]
        zeros = self.zeros.astype(np.float32)[:, :, None]
        out = (blocked * scales + zeros).reshape(self.packed.shape[0], padded)
        return out[:, : self.dim]

    def _dequantize_ids(self, token_ids: np.ndarray) -> np.ndarray:
        """Dequantize a subset of rows by token id (fast path uses cache)."""
        if self._embedding_table is not None:
            return self._embedding_table[token_ids]
        # Cold path: dequant unique ids only
        unique = np.unique(token_ids)
        packed = self.packed[unique]
        low = (packed & 0x0F).astype(np.float32)
        high = ((packed >> 4) & 0x0F).astype(np.float32)
        padded = packed.shape[1] * 2
        unpacked = np.empty((packed.shape[0], padded), dtype=np.float32)
        unpacked[:, 0::2] = low
        unpacked[:, 1::2] = high
        blocked = unpacked.reshape(packed.shape[0], self.num_blocks, self.block_size)
        scales = self.scales[unique].astype(np.float32)[:, :, None]
        zeros = self.zeros[unique].astype(np.float32)[:, :, None]
        deq = (blocked * scales + zeros).reshape(packed.shape[0], padded)[:, : self.dim]
        table = np.empty((self.vocab_size, self.dim), dtype=np.float32)
        table[unique] = deq
        self._embedding_table = table  # promote to cache
        return table[token_ids]

    # ---- SIF + PC fitting ----------------------------------------------

    def fit_idf(self, corpus_token_lists: Sequence[Sequence[int]]) -> "VortexEmbedV2":
        """Compute SIF (Smoothed Inverse Frequency) weights from the corpus.

        weight(t) = a / (a + p(t))   where p(t) = count(t) / total_tokens.

        Tokens that never appear in the corpus get weight 1 (no down-weight).
        Call once after tokenizing the corpus; reused for every encode.
        """
        flat = (np.concatenate(corpus_token_lists)
                if corpus_token_lists else np.empty(0, dtype=np.int64))
        total = max(int(flat.size), 1)
        counts = np.bincount(flat, minlength=self.vocab_size).astype(np.float64)
        p = counts / total
        denom = self.sif_a + p
        with np.errstate(divide="ignore", invalid="ignore"):
            weights = np.where(p > 0, self.sif_a / denom, 1.0)
        self._sif_weights = weights.astype(np.float32)
        return self

    def fit_pc(self, corpus_embeddings: np.ndarray, k: Optional[int] = None) -> "VortexEmbedV2":
        """Compute the top-``k`` principal components of the corpus embeddings.

        These directions capture the dominant "common-topic" axis and are
        removed from every chunk/query vector at encode time. SIF-style
        trick from Arora et al. 2017. ``k=8`` is the v2 default.
        """
        if k is None:
            k = self.pc_k
        if corpus_embeddings.size == 0 or k <= 0:
            return self
        x = corpus_embeddings.astype(np.float32)
        x = x - x.mean(axis=0, keepdims=True)
        try:
            _, _, vt = np.linalg.svd(x, full_matrices=False)
            pcs = vt[:k].astype(np.float32)
            pcs = pcs / (np.linalg.norm(pcs, axis=1, keepdims=True) + 1e-12)
            self._pc_directions = pcs
        except np.linalg.LinAlgError:
            self._pc_directions = None
        return self

    def _apply_pc(self, x: np.ndarray) -> np.ndarray:
        if self.sif_pc <= 0 or self._pc_directions is None:
            return x
        out = x
        for pc in self._pc_directions:
            proj = (out @ pc)[:, None] * pc[None, :]
            out = out - self.sif_pc * proj
        return out

    # ---- file-path binding ---------------------------------------------

    def set_file_paths(self, file_paths: Sequence[str]) -> "VortexEmbedV2":
        """Bind corpus file paths so encode() can prepend path headers.

        Also pre-classifies each chunk by extension so the search-time bias
        can be applied in a tight loop without per-query re-classification.
        """
        self._file_paths = list(file_paths)
        if file_paths is None:
            self._chunk_is_py = None
            self._chunk_is_md = None
            return self
        self._chunk_is_py = np.fromiter(
            (p.endswith(".py") for p in file_paths), dtype=bool, count=len(file_paths)
        )
        self._chunk_is_md = np.fromiter(
            (p.endswith(".md") for p in file_paths), dtype=bool, count=len(file_paths)
        )
        return self

    def _augment_texts(self, texts: Sequence[str]) -> List[str]:
        if self._file_paths is None or len(self._file_paths) != len(texts):
            return list(texts)
        out: List[str] = []
        for text, path in zip(texts, self._file_paths):
            header_tokens = _path_to_header_tokens(path)
            if not header_tokens or self.header_repeat <= 0:
                out.append(text)
                continue
            header = " ".join(header_tokens * self.header_repeat)
            out.append(f"{header}\n{text}")
        return out

    # ---- tokenization ----------------------------------------------------

    DEFAULT_MAX_CHARS_PER_TEXT = 50_000
    DEFAULT_MAX_TOKENS_PER_TEXT = 4096
    DEFAULT_MAX_TOKENS_PER_BATCH = 262_144

    def _tokenize_batch(self, texts: Sequence[str]) -> List[List[int]]:
        encoded = self.tokenizer.encode_batch(list(texts))
        return [
            [tid for tid in item.ids if 0 <= int(tid) < self.vocab_size]
            for item in encoded
        ]

    def _cap_inputs(self, texts: Sequence[str]) -> List[str]:
        cap = self.DEFAULT_MAX_CHARS_PER_TEXT
        if cap <= 0:
            return list(texts)
        out = []
        for t in texts:
            if len(t) <= cap:
                out.append(t)
            else:
                half = cap // 2
                out.append(t[:half] + t[-(cap - half):])
        return out

    def _cap_token_lists(self, token_lists: List[List[int]]) -> List[List[int]]:
        cap = self.DEFAULT_MAX_TOKENS_PER_TEXT
        if cap <= 0:
            return token_lists
        out = []
        for ids in token_lists:
            if len(ids) <= cap:
                out.append(ids)
            else:
                half = cap // 2
                out.append(ids[:half] + ids[-(cap - half):])
        return out

    @staticmethod
    def _normalize_inplace(x: np.ndarray) -> None:
        norms = np.linalg.norm(x, axis=1, keepdims=True)
        np.divide(x, np.maximum(norms, 1e-12), out=x)

    # ---- core encode -----------------------------------------------------

    def _encode_subbatch(
        self, token_lists: Sequence[Sequence[int]], *, normalize: bool
    ) -> np.ndarray:
        n = len(token_lists)
        flat = (np.concatenate(token_lists)
                if token_lists else np.empty(0, dtype=np.int64))
        if flat.size == 0:
            return np.zeros((n, self.dim), dtype=np.float32)

        token_embs = self._dequantize_ids(flat)

        if self._sif_weights is not None:
            w = self._sif_weights[flat].astype(np.float32)[:, None]
            token_embs = token_embs * w

        import torch
        ro = torch.from_numpy(
            np.repeat(np.arange(n, dtype=np.int64),
                      [len(ids) for ids in token_lists])
        )
        em = torch.from_numpy(np.ascontiguousarray(token_embs))
        sums = torch.zeros((n, self.dim), dtype=torch.float32)
        sums.index_add_(0, ro, em)

        if self._sif_weights is not None:
            w_flat = torch.from_numpy(self._sif_weights[flat])
            w_per_row = ro.bincount(minlength=n, weights=w_flat).clamp(min=1e-12)
        else:
            w_per_row = ro.bincount(minlength=n).clamp(min=1).to(torch.float32)

        embeddings = (sums / w_per_row.unsqueeze(1)).numpy()
        embeddings = self._apply_pc(embeddings)
        if normalize:
            self._normalize_inplace(embeddings)
        return embeddings

    def encode_batch(
        self,
        texts: Sequence[str],
        *,
        normalize: bool = True,
        max_tokens_per_text: Optional[int] = None,
        max_tokens_per_batch: Optional[int] = None,
        max_chars_per_text: Optional[int] = None,
    ) -> np.ndarray:
        """Encode a list of texts into L2-normalized ``(len, dim)`` embeddings.

        Path-header augmentation runs first if file paths were bound via
        :meth:`set_file_paths`. Token caps and sub-batching keep peak
        memory bounded on large corpora.
        """
        if not texts:
            return np.zeros((0, self.dim), dtype=np.float32)

        augmented = self._augment_texts(texts)
        capped = self._cap_inputs(augmented)
        token_lists = self._tokenize_batch(capped)
        token_lists = self._cap_token_lists(token_lists)

        cap_t = (self.DEFAULT_MAX_TOKENS_PER_TEXT
                 if max_tokens_per_text is None else int(max_tokens_per_text))
        cap_b = (self.DEFAULT_MAX_TOKENS_PER_BATCH
                 if max_tokens_per_batch is None else int(max_tokens_per_batch))
        _ = cap_t  # already applied above

        total_tokens = sum(len(ids) for ids in token_lists)
        if total_tokens == 0:
            return np.zeros((len(texts), self.dim), dtype=np.float32)

        # Single-pass fast path
        if total_tokens <= cap_b or len(texts) <= 1:
            return self._encode_subbatch(token_lists, normalize=normalize)

        # Multi-pass path: split so each sub-batch fits in cap_b tokens
        out = np.zeros((len(texts), self.dim), dtype=np.float32)
        sub: List[List[int]] = []
        sub_tokens = 0
        sub_start = 0
        for i, ids in enumerate(token_lists):
            if sub and (sub_tokens + len(ids) > cap_b):
                out[sub_start:i] = self._encode_subbatch(
                    token_lists[sub_start:i], normalize=False
                )
                sub_start = i
                sub = [ids]
                sub_tokens = len(ids)
            else:
                sub.append(ids)
                sub_tokens += len(ids)
        if sub:
            out[sub_start:] = self._encode_subbatch(
                token_lists[sub_start:], normalize=False
            )
        if normalize:
            self._normalize_inplace(out)
        return out

    def encode_batch_cached(
        self,
        texts: Sequence[str],
        *,
        normalize: bool = True,
        cache_path: Optional[Union[str, Path]] = None,
        **encode_kwargs,
    ) -> np.ndarray:
        """Encode with a SHA-1-keyed on-disk cache for fast re-indexing.

        Cache is keyed on the sorted SHA-1 of (texts, dim, tokenizer id).
        On a hit, returns a fresh array without re-running the encode
        pipeline. ``cache_path`` is a path prefix; the actual files are
        ``{cache_path}.npy`` (embeddings) and ``{cache_path}.json`` (meta).
        """
        if cache_path is None and self.cache_path is not None:
            cache_path = self.cache_path
        if cache_path is None:
            return self.encode_batch(texts, normalize=normalize, **encode_kwargs)
        cache_path = Path(cache_path)
        cache_path.parent.mkdir(parents=True, exist_ok=True)
        emb_path = cache_path.with_suffix(".npy")
        meta_path = cache_path.with_suffix(".json")
        import hashlib
        h = hashlib.sha1()
        h.update(f"{self.dim}|v2|{len(texts)}|".encode())
        for t in texts:
            h.update(t.encode("utf-8", errors="replace"))
            h.update(b"\x00")
        fp = h.hexdigest()
        if meta_path.exists() and emb_path.exists():
            try:
                meta = json.loads(meta_path.read_text())
                if meta.get("fingerprint") == fp and meta.get("dim") == self.dim:
                    cached = np.load(emb_path, mmap_mode=None)
                    if cached.shape == (len(texts), self.dim):
                        return cached.copy() if normalize else cached
            except Exception:
                pass
        emb = self.encode_batch(texts, normalize=normalize, **encode_kwargs)
        np.save(emb_path, emb.astype(np.float32))
        meta_path.write_text(json.dumps({"fingerprint": fp, "dim": self.dim, "n": len(texts)}))
        return emb

    def encode(self, texts: Union[str, Sequence[str]], *, normalize: bool = True) -> np.ndarray:
        """Encode one string or a list of strings.

        For a single string, returns a 1-D array of shape ``(dim,)``.
        For a list, returns a 2-D array of shape ``(len, dim)``.
        """
        if isinstance(texts, str):
            return self.encode_batch([texts], normalize=normalize)[0]
        return self.encode_batch(list(texts), normalize=normalize)

    # ---- search ---------------------------------------------------------

    def search(
        self,
        queries: np.ndarray,
        index: np.ndarray,
        top_k: int = 10,
        *,
        index_normalized: bool = True,
    ) -> Tuple[np.ndarray, np.ndarray]:
        """Cosine search with optional file-extension score bias.

        Returns ``(scores, indices)`` of shapes ``(Q, top_k)`` and
        ``(Q, top_k)``. Indices are row indices into ``index``.

        Set ``index_normalized=False`` to have the index L2-normalized
        in-place; otherwise it is assumed to be pre-normalized.
        """
        queries = np.asarray(queries, dtype=np.float32)
        index = np.asarray(index, dtype=np.float32)
        if queries.ndim == 1:
            queries = queries[None, :]
        if not index_normalized:
            index = index.copy()
            self._normalize_inplace(index)
        qn = queries.copy()
        self._normalize_inplace(qn)

        scores = qn @ index.T
        n_docs = scores.shape[1]
        k = min(int(top_k), n_docs)
        if k <= 0:
            return (np.empty((queries.shape[0], 0), dtype=np.float32),
                    np.empty((queries.shape[0], 0), dtype=np.int64))

        bias_pool = min(self.bias_top_k, n_docs)
        if bias_pool >= n_docs:
            order = np.argsort(-scores, axis=1)
        else:
            part = np.argpartition(-scores, bias_pool, axis=1)[:, :bias_pool]
            ps = np.take_along_axis(scores, part, axis=1)
            sub_order = np.argsort(-ps, axis=1)
            order = np.take_along_axis(part, sub_order, axis=1)

        # v2 search-time bias: vectorized score adjustment on the candidate
        # pool. Adds py_bonus to .py chunks and md_penalty to .md chunks in
        # the top-bias_pool per query, then a final argpartition/top-k.
        if self._chunk_is_py is not None or self._chunk_is_md is not None:
            biased = scores.copy()
            # Build a per-chunk additive bias vector once
            chunk_bias = np.zeros(scores.shape[1], dtype=np.float32)
            if self._chunk_is_py is not None:
                chunk_bias += np.where(self._chunk_is_py, self.py_bonus, 0.0)
            if self._chunk_is_md is not None:
                chunk_bias += np.where(self._chunk_is_md, self.md_penalty, 0.0)
            # Zero out bias for non-candidate docs (so they can never
            # outrank a candidate via the bias)
            mask = np.zeros(scores.shape[1], dtype=bool)
            for qi in range(scores.shape[0]):
                mask[order[qi]] = True
            chunk_bias = np.where(mask, chunk_bias, 0.0)
            biased += chunk_bias[None, :]
            scores = biased

        if k == n_docs:
            idx = np.argsort(-scores, axis=1)[:, :k]
        else:
            part = np.argpartition(-scores, kth=k, axis=1)[:, :k]
            ps = np.take_along_axis(scores, part, axis=1)
            order2 = np.argsort(-ps, axis=1)
            idx = np.take_along_axis(part, order2, axis=1)
        ordered_scores = np.take_along_axis(scores, idx, axis=1)
        return (ordered_scores.astype(np.float32, copy=False),
                idx.astype(np.int64, copy=False))