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 | class DuckDBVectorStore(BasePydanticVectorStore):
"""DuckDB vector store.
In this vector store, embeddings are stored within a DuckDB database.
During query time, the index uses DuckDB to query for the top
k most similar nodes.
Examples:
`pip install llama-index-vector-stores-duckdb`
```python
from llama_index.vector_stores.duckdb import DuckDBVectorStore
# in-memory
vector_store = DuckDBVectorStore()
# persist to disk
vector_store = DuckDBVectorStore("pg.duckdb", persist_dir="./persist/")
```
"""
stores_text: bool = True
flat_metadata: bool = True
database_name: Optional[str]
table_name: Optional[str]
# schema_name: Optional[str] # TODO: support schema name
embed_dim: Optional[int]
# hybrid_search: Optional[bool] # TODO: support hybrid search
text_search_config: Optional[dict]
persist_dir: Optional[str]
_conn: Any = PrivateAttr()
_is_initialized: bool = PrivateAttr(default=False)
_database_path: Optional[str] = PrivateAttr()
def __init__(
self,
database_name: Optional[str] = ":memory:",
table_name: Optional[str] = "documents",
# schema_name: Optional[str] = "main",
embed_dim: Optional[int] = None,
# hybrid_search: Optional[bool] = False,
# https://duckdb.org/docs/extensions/full_text_search
text_search_config: Optional[dict] = {
"stemmer": "english",
"stopwords": "english",
"ignore": "(\\.|[^a-z])+",
"strip_accents": True,
"lower": True,
"overwrite": False,
},
persist_dir: Optional[str] = "./storage",
**kwargs: Any,
) -> None:
"""Init params."""
try:
import duckdb
except ImportError:
raise ImportError(import_err_msg)
self._is_initialized = False
if database_name == ":memory:":
_home_dir = os.path.expanduser("~")
self._conn = duckdb.connect(database_name)
self._conn.execute(f"SET home_directory='{_home_dir}';")
self._conn.install_extension("json")
self._conn.load_extension("json")
self._conn.install_extension("fts")
self._conn.load_extension("fts")
else:
# check if persist dir exists
if not os.path.exists(persist_dir):
os.makedirs(persist_dir)
self._database_path = os.path.join(persist_dir, database_name)
with DuckDBLocalContext(self._database_path) as _conn:
pass
self._conn = None
super().__init__(
database_name=database_name,
table_name=table_name,
# schema_name=schema_name,
embed_dim=embed_dim,
# hybrid_search=hybrid_search,
text_search_config=text_search_config,
persist_dir=persist_dir,
)
@classmethod
def from_local(
cls, database_path: str, table_name: str = "documents"
) -> "DuckDBVectorStore":
"""Load a DuckDB vector store from a local file."""
with DuckDBLocalContext(database_path) as _conn:
try:
_table_info = _conn.execute(f"SHOW {table_name};").fetchall()
except Exception as e:
raise ValueError(f"Index table {table_name} not found in the database.")
# Not testing for the column type similarity only testing for the column names.
_std = {"text", "node_id", "embedding", "metadata_"}
_ti = {_i[0] for _i in _table_info}
if _std != _ti:
raise ValueError(
f"Index table {table_name} does not have the correct schema."
)
_cls = cls(
database_name=os.path.basename(database_path),
table_name=table_name,
persist_dir=os.path.dirname(database_path),
)
_cls._is_initialized = True
return _cls
@classmethod
def from_params(
cls,
database_name: Optional[str] = ":memory:",
table_name: Optional[str] = "documents",
# schema_name: Optional[str] = "main",
embed_dim: Optional[int] = None,
# hybrid_search: Optional[bool] = False,
text_search_config: Optional[dict] = {
"stemmer": "english",
"stopwords": "english",
"ignore": "(\\.|[^a-z])+",
"strip_accents": True,
"lower": True,
"overwrite": False,
},
persist_dir: Optional[str] = "./storage",
**kwargs: Any,
) -> "DuckDBVectorStore":
return cls(
database_name=database_name,
table_name=table_name,
# schema_name=schema_name,
embed_dim=embed_dim,
# hybrid_search=hybrid_search,
text_search_config=text_search_config,
persist_dir=persist_dir,
**kwargs,
)
@classmethod
def class_name(cls) -> str:
return "DuckDBVectorStore"
@property
def client(self) -> Any:
"""Return client."""
return self._conn
def _initialize(self) -> None:
if not self._is_initialized:
# TODO: schema.table also.
# Check if table and type is present
# if not, create table
if self.embed_dim is None:
_query = f"""
CREATE TABLE {self.table_name} (
node_id VARCHAR,
text TEXT,
embedding FLOAT[],
metadata_ JSON
);
"""
else:
_query = f"""
CREATE TABLE {self.table_name} (
node_id VARCHAR,
text TEXT,
embedding FLOAT[{self.embed_dim}],
metadata_ JSON
);
"""
if self.database_name == ":memory:":
self._conn.execute(_query)
else:
with DuckDBLocalContext(self._database_path) as _conn:
_conn.execute(_query)
self._is_initialized = True
def _node_to_table_row(self, node: BaseNode) -> Any:
return (
node.node_id,
node.get_content(metadata_mode=MetadataMode.NONE),
node.get_embedding(),
node_to_metadata_dict(
node,
remove_text=True,
flat_metadata=self.flat_metadata,
),
)
def add(self, nodes: List[BaseNode], **add_kwargs: Any) -> List[str]:
"""Add nodes to index.
Args:
nodes: List[BaseNode]: list of nodes with embeddings
"""
self._initialize()
ids = []
if self.database_name == ":memory:":
_table = self._conn.table(self.table_name)
for node in nodes:
ids.append(node.node_id)
_row = self._node_to_table_row(node)
_table.insert(_row)
else:
with DuckDBLocalContext(self._database_path) as _conn:
_table = _conn.table(self.table_name)
for node in nodes:
ids.append(node.node_id)
_row = self._node_to_table_row(node)
_table.insert(_row)
return ids
def delete(self, ref_doc_id: str, **delete_kwargs: Any) -> None:
"""
Delete nodes using with ref_doc_id.
Args:
ref_doc_id (str): The doc_id of the document to delete.
"""
_ddb_query = f"""
DELETE FROM {self.table_name}
WHERE json_extract_string(metadata_, '$.ref_doc_id') = '{ref_doc_id}';
"""
if self.database_name == ":memory:":
self._conn.execute(_ddb_query)
else:
with DuckDBLocalContext(self._database_path) as _conn:
_conn.execute(_ddb_query)
@staticmethod
def _build_metadata_filter_condition(
standard_filters: MetadataFilters,
) -> dict:
"""Translate standard metadata filters to DuckDB SQL specification."""
filters_list = []
# condition = standard_filters.condition or "and" ## and/or as strings.
condition = "AND"
_filters_condition_list = []
for filter in standard_filters.filters:
if filter.operator:
if filter.operator in [
"<",
">",
"<=",
">=",
"<>",
"!=",
]:
filters_list.append((filter.key, filter.operator, filter.value))
elif filter.operator in ["=="]:
filters_list.append((filter.key, "=", filter.value))
else:
raise ValueError(
f"Filter operator {filter.operator} not supported."
)
else:
filters_list.append((filter.key, "=", filter.value))
for _fc in filters_list:
if isinstance(_fc[2], str):
_filters_condition_list.append(
f"json_extract_string(metadata_, '$.{_fc[0]}') {_fc[1]} '{_fc[2]}'"
)
else:
_filters_condition_list.append(
f"json_extract(metadata_, '$.{_fc[0]}') {_fc[1]} {_fc[2]}"
)
return f" {condition} ".join(_filters_condition_list)
def query(self, query: VectorStoreQuery, **kwargs: Any) -> VectorStoreQueryResult:
"""Query index for top k most similar nodes.
Args:
query.query_embedding (List[float]): query embedding
query.similarity_top_k (int): top k most similar nodes
"""
nodes = []
similarities = []
ids = []
if query.filters is not None:
# TODO: results from the metadata filter query
_filter_string = self._build_metadata_filter_condition(query.filters)
_ddb_query = f"""
SELECT node_id, text, embedding, metadata_, score
FROM (
SELECT *, list_cosine_similarity(embedding, {query.query_embedding}) AS score
FROM {self.table_name}
WHERE {_filter_string}
) sq
WHERE score IS NOT NULL
ORDER BY score DESC LIMIT {query.similarity_top_k};
"""
else:
_ddb_query = f"""
SELECT node_id, text, embedding, metadata_, score
FROM (
SELECT *, list_cosine_similarity(embedding, {query.query_embedding}) AS score
FROM {self.table_name}
) sq
WHERE score IS NOT NULL
ORDER BY score DESC LIMIT {query.similarity_top_k};
"""
if self.database_name == ":memory:":
_final_results = self._conn.execute(_ddb_query).fetchall()
else:
with DuckDBLocalContext(self._database_path) as _conn:
_final_results = _conn.execute(_ddb_query).fetchall()
for _row in _final_results:
node = TextNode(
id_=_row[0],
text=_row[1],
embedding=_row[2],
metadata=json.loads(_row[3]),
)
nodes.append(node)
similarities.append(_row[4])
ids.append(_row[0])
return VectorStoreQueryResult(nodes=nodes, similarities=similarities, ids=ids)
|