File size: 10,190 Bytes
935c8f7 610fb0e 935c8f7 | 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 | """Persistent chunk and embedding artifact loading for document reuse."""
from __future__ import annotations
from dataclasses import dataclass
from typing import Any, Callable, Optional, Sequence
import numpy as np
from sqlalchemy import select
EMBEDDING_DTYPE = "float32"
class ArtifactValidationError(ValueError):
"""Raised when persisted chunk ordering cannot be trusted."""
@dataclass(frozen=True)
class StoredDocumentArtifacts:
document_id: str
chunks: list[dict[str, Any]]
embedding_matrix: Optional[np.ndarray]
needs_embedding_backfill: bool
def serialize_embedding(embedding: Any) -> tuple[bytes, int, str]:
vector = np.asarray(embedding, dtype=np.float32)
if vector.ndim != 1 or vector.size == 0:
raise ArtifactValidationError("A chunk embedding must be a non-empty one-dimensional vector.")
contiguous_vector = np.ascontiguousarray(vector, dtype=np.float32)
return contiguous_vector.tobytes(), int(contiguous_vector.shape[0]), EMBEDDING_DTYPE
def deserialize_embedding(blob: Any, dimension: Any, dtype: Any) -> np.ndarray:
if blob is None or dimension is None or dtype != EMBEDDING_DTYPE:
raise ArtifactValidationError("A stored chunk embedding is missing required metadata.")
try:
parsed_dimension = int(dimension)
except (TypeError, ValueError) as exc:
raise ArtifactValidationError("A stored embedding dimension is invalid.") from exc
if parsed_dimension <= 0:
raise ArtifactValidationError("A stored embedding dimension must be positive.")
vector = np.frombuffer(bytes(blob), dtype=np.float32)
if vector.size != parsed_dimension:
raise ArtifactValidationError("A stored embedding byte length does not match its dimension.")
return np.ascontiguousarray(vector, dtype=np.float32)
def embedding_matrix_from_chunk_rows(chunk_rows: Sequence[Any]) -> np.ndarray:
if not chunk_rows:
raise ArtifactValidationError("A persisted document has no chunks.")
expected_indices = list(range(len(chunk_rows)))
actual_indices = [int(row.chunk_index) for row in chunk_rows]
if actual_indices != expected_indices:
raise ArtifactValidationError("Persisted chunks are not in contiguous chunk_index order.")
vectors = [
deserialize_embedding(row.embedding_blob, row.embedding_dimension, row.embedding_dtype)
for row in chunk_rows
]
dimensions = {int(vector.shape[0]) for vector in vectors}
if len(dimensions) != 1:
raise ArtifactValidationError("Persisted chunk embedding dimensions do not match.")
matrix = np.vstack(vectors).astype(np.float32, copy=False)
matrix = np.ascontiguousarray(matrix, dtype=np.float32)
if matrix.ndim != 2 or matrix.shape[0] != len(chunk_rows):
raise ArtifactValidationError("Persisted embedding count does not match chunk count.")
return matrix
def load_owned_document_query_artifacts(
*, user_id: str, document_id: str, embedding_model: str, embedding_format: str,
) -> StoredDocumentArtifacts:
"""Load ready, owned chunks and stored vectors without embedding backfill."""
from persistence.db import SessionLocal
from persistence.models import Chunk
from persistence.ownership import require_owned_document
with SessionLocal() as session:
document = require_owned_document(session, user_id=user_id, document_id=document_id)
if document.status != "ready":
raise ArtifactValidationError("Document is not ready for querying.")
if document.embedding_model != embedding_model or document.embedding_format != embedding_format:
raise ArtifactValidationError("Stored embeddings do not match the configured model.")
rows = session.execute(
select(Chunk)
.where(Chunk.user_id == user_id, Chunk.document_id == document_id)
.order_by(Chunk.chunk_index.asc())
).scalars().all()
matrix = embedding_matrix_from_chunk_rows(rows)
return StoredDocumentArtifacts(
document_id=document_id,
chunks=[
{"text": row.text, "page": int(row.page_number), "chunk_id": int(row.chunk_id)}
for row in rows
],
embedding_matrix=matrix,
needs_embedding_backfill=False,
)
def load_document_artifacts(
*,
user_id: str,
source_hash: str,
embedding_model: str,
embedding_format: str,
) -> Optional[StoredDocumentArtifacts]:
from persistence.db import SessionLocal
from persistence.models import Chunk, Document
with SessionLocal() as session:
document = session.execute(
select(Document)
.where(Document.user_id == user_id, Document.source_hash == source_hash)
.order_by(Document.created_at.desc())
).scalars().first()
if document is None:
return None
chunk_rows = session.execute(
select(Chunk)
.where(Chunk.user_id == user_id, Chunk.document_id == document.id)
.order_by(Chunk.chunk_index.asc())
).scalars().all()
expected_indices = list(range(len(chunk_rows)))
actual_indices = [int(row.chunk_index) for row in chunk_rows]
if actual_indices != expected_indices:
raise ArtifactValidationError("Persisted chunks are not in contiguous chunk_index order.")
chunks = [
{
"text": row.text,
"page": int(row.page_number),
"chunk_id": int(row.chunk_id),
}
for row in chunk_rows
]
embedding_matrix: Optional[np.ndarray] = None
configuration_matches = (
document.embedding_model == embedding_model
and document.embedding_format == embedding_format
)
if configuration_matches and chunk_rows:
try:
embedding_matrix = embedding_matrix_from_chunk_rows(chunk_rows)
except ArtifactValidationError:
embedding_matrix = None
return StoredDocumentArtifacts(
document_id=document.id,
chunks=chunks,
embedding_matrix=embedding_matrix,
needs_embedding_backfill=embedding_matrix is None,
)
def load_or_backfill_document_embeddings(
*,
user_id: str,
document_id: str,
source_hash: str,
embedding_model: str,
embedding_format: str,
embedding_generator: Callable[[list[dict[str, Any]]], np.ndarray],
) -> StoredDocumentArtifacts:
"""Reload and, when still necessary, repair embeddings under a document lock."""
from persistence.db import SessionLocal, mysql_document_lock
from persistence.models import Chunk, Document
with SessionLocal() as session:
try:
with mysql_document_lock(
session,
user_id=user_id,
document_key_parts=(source_hash,),
):
document = session.execute(
select(Document).where(
Document.id == document_id,
Document.user_id == user_id,
Document.source_hash == source_hash,
)
).scalar_one_or_none()
if document is None:
raise ArtifactValidationError("Document disappeared before embedding backfill.")
chunk_rows = session.execute(
select(Chunk)
.where(Chunk.user_id == user_id, Chunk.document_id == document_id)
.order_by(Chunk.chunk_index.asc())
).scalars().all()
expected_indices = list(range(len(chunk_rows)))
if [int(row.chunk_index) for row in chunk_rows] != expected_indices:
raise ArtifactValidationError("Persisted chunks are not in contiguous chunk_index order.")
if not chunk_rows:
raise ArtifactValidationError("A persisted document has no chunks.")
chunks = [
{
"text": row.text,
"page": int(row.page_number),
"chunk_id": int(row.chunk_id),
}
for row in chunk_rows
]
matrix: Optional[np.ndarray] = None
configuration_matches = (
document.embedding_model == embedding_model
and document.embedding_format == embedding_format
)
if configuration_matches:
try:
matrix = embedding_matrix_from_chunk_rows(chunk_rows)
except ArtifactValidationError:
matrix = None
if matrix is None:
matrix = np.ascontiguousarray(embedding_generator(chunks), dtype=np.float32)
if matrix.ndim != 2 or matrix.shape[0] != len(chunk_rows) or matrix.shape[1] == 0:
raise ArtifactValidationError("Embedding backfill count does not match chunk count.")
for row, vector in zip(chunk_rows, matrix):
blob, dimension, dtype = serialize_embedding(vector)
row.embedding_blob = blob
row.embedding_dimension = dimension
row.embedding_dtype = dtype
document.embedding_model = embedding_model
document.embedding_format = embedding_format
session.flush()
persisted_document_id = str(document.id)
session.commit()
else:
persisted_document_id = str(document.id)
result = StoredDocumentArtifacts(
document_id=persisted_document_id,
chunks=chunks,
embedding_matrix=matrix,
needs_embedding_backfill=False,
)
return result
except Exception:
session.rollback()
raise
|