File size: 20,963 Bytes
129cd69
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Module contains logic for indexing documents into vector stores."""
from __future__ import annotations

import hashlib
import json
import uuid
from itertools import islice
from typing import (
    Any,
    AsyncIterable,
    AsyncIterator,
    Callable,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Optional,
    Sequence,
    Set,
    TypedDict,
    TypeVar,
    Union,
    cast,
)

from langchain_core.documents import Document
from langchain_core.pydantic_v1 import root_validator
from langchain_core.vectorstores import VectorStore

from langchain.document_loaders.base import BaseLoader
from langchain.indexes.base import NAMESPACE_UUID, RecordManager

T = TypeVar("T")


def _hash_string_to_uuid(input_string: str) -> uuid.UUID:
    """Hashes a string and returns the corresponding UUID."""
    hash_value = hashlib.sha1(input_string.encode("utf-8")).hexdigest()
    return uuid.uuid5(NAMESPACE_UUID, hash_value)


def _hash_nested_dict_to_uuid(data: dict[Any, Any]) -> uuid.UUID:
    """Hashes a nested dictionary and returns the corresponding UUID."""
    serialized_data = json.dumps(data, sort_keys=True)
    hash_value = hashlib.sha1(serialized_data.encode("utf-8")).hexdigest()
    return uuid.uuid5(NAMESPACE_UUID, hash_value)


class _HashedDocument(Document):
    """A hashed document with a unique ID."""

    uid: str
    hash_: str
    """The hash of the document including content and metadata."""
    content_hash: str
    """The hash of the document content."""
    metadata_hash: str
    """The hash of the document metadata."""

    @root_validator(pre=True)
    def calculate_hashes(cls, values: Dict[str, Any]) -> Dict[str, Any]:
        """Root validator to calculate content and metadata hash."""
        content = values.get("page_content", "")
        metadata = values.get("metadata", {})

        forbidden_keys = ("hash_", "content_hash", "metadata_hash")

        for key in forbidden_keys:
            if key in metadata:
                raise ValueError(
                    f"Metadata cannot contain key {key} as it "
                    f"is reserved for internal use."
                )

        content_hash = str(_hash_string_to_uuid(content))

        try:
            metadata_hash = str(_hash_nested_dict_to_uuid(metadata))
        except Exception as e:
            raise ValueError(
                f"Failed to hash metadata: {e}. "
                f"Please use a dict that can be serialized using json."
            )

        values["content_hash"] = content_hash
        values["metadata_hash"] = metadata_hash
        values["hash_"] = str(_hash_string_to_uuid(content_hash + metadata_hash))

        _uid = values.get("uid", None)

        if _uid is None:
            values["uid"] = values["hash_"]
        return values

    def to_document(self) -> Document:
        """Return a Document object."""
        return Document(
            page_content=self.page_content,
            metadata=self.metadata,
        )

    @classmethod
    def from_document(
        cls, document: Document, *, uid: Optional[str] = None
    ) -> _HashedDocument:
        """Create a HashedDocument from a Document."""
        return cls(
            uid=uid,
            page_content=document.page_content,
            metadata=document.metadata,
        )


def _batch(size: int, iterable: Iterable[T]) -> Iterator[List[T]]:
    """Utility batching function."""
    it = iter(iterable)
    while True:
        chunk = list(islice(it, size))
        if not chunk:
            return
        yield chunk


async def _abatch(size: int, iterable: AsyncIterable[T]) -> AsyncIterator[List[T]]:
    """Utility batching function."""
    batch: List[T] = []
    async for element in iterable:
        if len(batch) < size:
            batch.append(element)

        if len(batch) >= size:
            yield batch
            batch = []

    if batch:
        yield batch


def _get_source_id_assigner(
    source_id_key: Union[str, Callable[[Document], str], None],
) -> Callable[[Document], Union[str, None]]:
    """Get the source id from the document."""
    if source_id_key is None:
        return lambda doc: None
    elif isinstance(source_id_key, str):
        return lambda doc: doc.metadata[source_id_key]
    elif callable(source_id_key):
        return source_id_key
    else:
        raise ValueError(
            f"source_id_key should be either None, a string or a callable. "
            f"Got {source_id_key} of type {type(source_id_key)}."
        )


def _deduplicate_in_order(
    hashed_documents: Iterable[_HashedDocument],
) -> Iterator[_HashedDocument]:
    """Deduplicate a list of hashed documents while preserving order."""
    seen: Set[str] = set()

    for hashed_doc in hashed_documents:
        if hashed_doc.hash_ not in seen:
            seen.add(hashed_doc.hash_)
            yield hashed_doc


# PUBLIC API


class IndexingResult(TypedDict):
    """Return a detailed a breakdown of the result of the indexing operation."""

    num_added: int
    """Number of added documents."""
    num_updated: int
    """Number of updated documents because they were not up to date."""
    num_deleted: int
    """Number of deleted documents."""
    num_skipped: int
    """Number of skipped documents because they were already up to date."""


def index(
    docs_source: Union[BaseLoader, Iterable[Document]],
    record_manager: RecordManager,
    vector_store: VectorStore,
    *,
    batch_size: int = 100,
    cleanup: Literal["incremental", "full", None] = None,
    source_id_key: Union[str, Callable[[Document], str], None] = None,
    cleanup_batch_size: int = 1_000,
) -> IndexingResult:
    """Index data from the loader into the vector store.

    Indexing functionality uses a manager to keep track of which documents
    are in the vector store.

    This allows us to keep track of which documents were updated, and which
    documents were deleted, which documents should be skipped.

    For the time being, documents are indexed using their hashes, and users
     are not able to specify the uid of the document.

    IMPORTANT:
       if auto_cleanup is set to True, the loader should be returning
       the entire dataset, and not just a subset of the dataset.
       Otherwise, the auto_cleanup will remove documents that it is not
       supposed to.

    Args:
        docs_source: Data loader or iterable of documents to index.
        record_manager: Timestamped set to keep track of which documents were
                         updated.
        vector_store: Vector store to index the documents into.
        batch_size: Batch size to use when indexing.
        cleanup: How to handle clean up of documents.
            - Incremental: Cleans up all documents that haven't been updated AND
                           that are associated with source ids that were seen
                           during indexing.
                           Clean up is done continuously during indexing helping
                           to minimize the probability of users seeing duplicated
                           content.
            - Full: Delete all documents that haven to been returned by the loader.
                    Clean up runs after all documents have been indexed.
                    This means that users may see duplicated content during indexing.
            - None: Do not delete any documents.
        source_id_key: Optional key that helps identify the original source
            of the document.
        cleanup_batch_size: Batch size to use when cleaning up documents.

    Returns:
        Indexing result which contains information about how many documents
        were added, updated, deleted, or skipped.
    """
    if cleanup not in {"incremental", "full", None}:
        raise ValueError(
            f"cleanup should be one of 'incremental', 'full' or None. "
            f"Got {cleanup}."
        )

    if cleanup == "incremental" and source_id_key is None:
        raise ValueError("Source id key is required when cleanup mode is incremental.")

    # Check that the Vectorstore has required methods implemented
    methods = ["delete", "add_documents"]

    for method in methods:
        if not hasattr(vector_store, method):
            raise ValueError(
                f"Vectorstore {vector_store} does not have required method {method}"
            )

    if type(vector_store).delete == VectorStore.delete:
        # Checking if the vectorstore has overridden the default delete method
        # implementation which just raises a NotImplementedError
        raise ValueError("Vectorstore has not implemented the delete method")

    if isinstance(docs_source, BaseLoader):
        try:
            doc_iterator = docs_source.lazy_load()
        except NotImplementedError:
            doc_iterator = iter(docs_source.load())
    else:
        doc_iterator = iter(docs_source)

    source_id_assigner = _get_source_id_assigner(source_id_key)

    # Mark when the update started.
    index_start_dt = record_manager.get_time()
    num_added = 0
    num_skipped = 0
    num_updated = 0
    num_deleted = 0

    for doc_batch in _batch(batch_size, doc_iterator):
        hashed_docs = list(
            _deduplicate_in_order(
                [_HashedDocument.from_document(doc) for doc in doc_batch]
            )
        )

        source_ids: Sequence[Optional[str]] = [
            source_id_assigner(doc) for doc in hashed_docs
        ]

        if cleanup == "incremental":
            # If the cleanup mode is incremental, source ids are required.
            for source_id, hashed_doc in zip(source_ids, hashed_docs):
                if source_id is None:
                    raise ValueError(
                        "Source ids are required when cleanup mode is incremental. "
                        f"Document that starts with "
                        f"content: {hashed_doc.page_content[:100]} was not assigned "
                        f"as source id."
                    )
            # source ids cannot be None after for loop above.
            source_ids = cast(Sequence[str], source_ids)  # type: ignore[assignment]

        exists_batch = record_manager.exists([doc.uid for doc in hashed_docs])

        # Filter out documents that already exist in the record store.
        uids = []
        docs_to_index = []
        for hashed_doc, doc_exists in zip(hashed_docs, exists_batch):
            if doc_exists:
                # Must be updated to refresh timestamp.
                record_manager.update([hashed_doc.uid], time_at_least=index_start_dt)
                num_skipped += 1
                continue
            uids.append(hashed_doc.uid)
            docs_to_index.append(hashed_doc.to_document())

        # Be pessimistic and assume that all vector store write will fail.
        # First write to vector store
        if docs_to_index:
            vector_store.add_documents(docs_to_index, ids=uids)
            num_added += len(docs_to_index)

        # And only then update the record store.
        # Update ALL records, even if they already exist since we want to refresh
        # their timestamp.
        record_manager.update(
            [doc.uid for doc in hashed_docs],
            group_ids=source_ids,
            time_at_least=index_start_dt,
        )

        # If source IDs are provided, we can do the deletion incrementally!
        if cleanup == "incremental":
            # Get the uids of the documents that were not returned by the loader.

            # mypy isn't good enough to determine that source ids cannot be None
            # here due to a check that's happening above, so we check again.
            for source_id in source_ids:
                if source_id is None:
                    raise AssertionError("Source ids cannot be None here.")

            _source_ids = cast(Sequence[str], source_ids)

            uids_to_delete = record_manager.list_keys(
                group_ids=_source_ids, before=index_start_dt
            )
            if uids_to_delete:
                # Then delete from vector store.
                vector_store.delete(uids_to_delete)
                # First delete from record store.
                record_manager.delete_keys(uids_to_delete)
                num_deleted += len(uids_to_delete)

    if cleanup == "full":
        while uids_to_delete := record_manager.list_keys(
            before=index_start_dt, limit=cleanup_batch_size
        ):
            # First delete from record store.
            vector_store.delete(uids_to_delete)
            # Then delete from record manager.
            record_manager.delete_keys(uids_to_delete)
            num_deleted += len(uids_to_delete)

    return {
        "num_added": num_added,
        "num_updated": num_updated,
        "num_skipped": num_skipped,
        "num_deleted": num_deleted,
    }


# Define an asynchronous generator function
async def _to_async_iterator(iterator: Iterable[T]) -> AsyncIterator[T]:
    """Convert an iterable to an async iterator."""
    for item in iterator:
        yield item


async def aindex(
    docs_source: Union[Iterable[Document], AsyncIterator[Document]],
    record_manager: RecordManager,
    vector_store: VectorStore,
    *,
    batch_size: int = 100,
    cleanup: Literal["incremental", "full", None] = None,
    source_id_key: Union[str, Callable[[Document], str], None] = None,
    cleanup_batch_size: int = 1_000,
) -> IndexingResult:
    """Index data from the loader into the vector store.

    Indexing functionality uses a manager to keep track of which documents
    are in the vector store.

    This allows us to keep track of which documents were updated, and which
    documents were deleted, which documents should be skipped.

    For the time being, documents are indexed using their hashes, and users
     are not able to specify the uid of the document.

    IMPORTANT:
       if auto_cleanup is set to True, the loader should be returning
       the entire dataset, and not just a subset of the dataset.
       Otherwise, the auto_cleanup will remove documents that it is not
       supposed to.

    Args:
        docs_source: Data loader or iterable of documents to index.
        record_manager: Timestamped set to keep track of which documents were
                         updated.
        vector_store: Vector store to index the documents into.
        batch_size: Batch size to use when indexing.
        cleanup: How to handle clean up of documents.
            - Incremental: Cleans up all documents that haven't been updated AND
                           that are associated with source ids that were seen
                           during indexing.
                           Clean up is done continuously during indexing helping
                           to minimize the probability of users seeing duplicated
                           content.
            - Full: Delete all documents that haven to been returned by the loader.
                    Clean up runs after all documents have been indexed.
                    This means that users may see duplicated content during indexing.
            - None: Do not delete any documents.
        source_id_key: Optional key that helps identify the original source
            of the document.
        cleanup_batch_size: Batch size to use when cleaning up documents.

    Returns:
        Indexing result which contains information about how many documents
        were added, updated, deleted, or skipped.
    """

    if cleanup not in {"incremental", "full", None}:
        raise ValueError(
            f"cleanup should be one of 'incremental', 'full' or None. "
            f"Got {cleanup}."
        )

    if cleanup == "incremental" and source_id_key is None:
        raise ValueError("Source id key is required when cleanup mode is incremental.")

    # Check that the Vectorstore has required methods implemented
    methods = ["adelete", "aadd_documents"]

    for method in methods:
        if not hasattr(vector_store, method):
            raise ValueError(
                f"Vectorstore {vector_store} does not have required method {method}"
            )

    if type(vector_store).adelete == VectorStore.adelete:
        # Checking if the vectorstore has overridden the default delete method
        # implementation which just raises a NotImplementedError
        raise ValueError("Vectorstore has not implemented the delete method")

    if isinstance(docs_source, BaseLoader):
        raise NotImplementedError(
            "Not supported yet. Please pass an async iterator of documents."
        )
    async_doc_iterator: AsyncIterator[Document]

    if hasattr(docs_source, "__aiter__"):
        async_doc_iterator = docs_source  # type: ignore[assignment]
    else:
        async_doc_iterator = _to_async_iterator(docs_source)

    source_id_assigner = _get_source_id_assigner(source_id_key)

    # Mark when the update started.
    index_start_dt = await record_manager.aget_time()
    num_added = 0
    num_skipped = 0
    num_updated = 0
    num_deleted = 0

    async for doc_batch in _abatch(batch_size, async_doc_iterator):
        hashed_docs = list(
            _deduplicate_in_order(
                [_HashedDocument.from_document(doc) for doc in doc_batch]
            )
        )

        source_ids: Sequence[Optional[str]] = [
            source_id_assigner(doc) for doc in hashed_docs
        ]

        if cleanup == "incremental":
            # If the cleanup mode is incremental, source ids are required.
            for source_id, hashed_doc in zip(source_ids, hashed_docs):
                if source_id is None:
                    raise ValueError(
                        "Source ids are required when cleanup mode is incremental. "
                        f"Document that starts with "
                        f"content: {hashed_doc.page_content[:100]} was not assigned "
                        f"as source id."
                    )
            # source ids cannot be None after for loop above.
            source_ids = cast(Sequence[str], source_ids)

        exists_batch = await record_manager.aexists([doc.uid for doc in hashed_docs])

        # Filter out documents that already exist in the record store.
        uids: list[str] = []
        docs_to_index: list[Document] = []

        for hashed_doc, doc_exists in zip(hashed_docs, exists_batch):
            if doc_exists:
                # Must be updated to refresh timestamp.
                await record_manager.aupdate(
                    [hashed_doc.uid], time_at_least=index_start_dt
                )
                num_skipped += 1
                continue
            uids.append(hashed_doc.uid)
            docs_to_index.append(hashed_doc.to_document())

        # Be pessimistic and assume that all vector store write will fail.
        # First write to vector store
        if docs_to_index:
            await vector_store.aadd_documents(docs_to_index, ids=uids)
            num_added += len(docs_to_index)

        # And only then update the record store.
        # Update ALL records, even if they already exist since we want to refresh
        # their timestamp.
        await record_manager.aupdate(
            [doc.uid for doc in hashed_docs],
            group_ids=source_ids,
            time_at_least=index_start_dt,
        )

        # If source IDs are provided, we can do the deletion incrementally!

        if cleanup == "incremental":
            # Get the uids of the documents that were not returned by the loader.

            # mypy isn't good enough to determine that source ids cannot be None
            # here due to a check that's happening above, so we check again.
            for source_id in source_ids:
                if source_id is None:
                    raise AssertionError("Source ids cannot be None here.")

            _source_ids = cast(Sequence[str], source_ids)

            uids_to_delete = await record_manager.alist_keys(
                group_ids=_source_ids, before=index_start_dt
            )
            if uids_to_delete:
                # Then delete from vector store.
                await vector_store.adelete(uids_to_delete)
                # First delete from record store.
                await record_manager.adelete_keys(uids_to_delete)
                num_deleted += len(uids_to_delete)

    if cleanup == "full":
        while uids_to_delete := await record_manager.alist_keys(
            before=index_start_dt, limit=cleanup_batch_size
        ):
            # First delete from record store.
            await vector_store.adelete(uids_to_delete)
            # Then delete from record manager.
            await record_manager.adelete_keys(uids_to_delete)
            num_deleted += len(uids_to_delete)

    return {
        "num_added": num_added,
        "num_updated": num_updated,
        "num_skipped": num_skipped,
        "num_deleted": num_deleted,
    }