langchain/libs/community/langchain_community/vectorstores/pgvector.py
Andreas Motl 17e42bbd18
community[patch]: pgvector: Slight refactoring to make code a bit more reusable (#16243)
- **Description:** Improve [pgvector vector store
adapter](https://github.com/langchain-ai/langchain/blob/v0.1.1/libs/community/langchain_community/vectorstores/pgvector.py)
to make it reusable by adapters deriving from that.
  - **Issue:** NA
  - **Dependencies:** NA
  - **References:** https://github.com/crate-workbench/langchain/pull/1
  - **Addressed to:** @eyurtsev, @cbornet


Hi from the CrateDB team,

first of all, thanks a stack for conceiving and maintaining LangChain.
We are currently [preparing a
patch](https://github.com/crate-workbench/langchain/pull/1) for adding
[CrateDB](https://github.com/crate/crate) to the list of community
adapters.

Because CrateDB aims to be compatible with PostgreSQL to some degree,
the vector store subsystem in LangChain derives functionality from the
corresponding implementation for pgvector.

Therefore, in order to make the implementation more reusable, we needed
to rename the private methods `__from` and `__query_collection` to the
less private counterparts `_from` and `_query_collection`, so they can
be overwritten, in order to unlock other adapters deriving from
[pgvector](https://github.com/langchain-ai/langchain/blob/v0.1.1/libs/community/langchain_community/vectorstores/pgvector.py).

With kind regards,
Andreas.
2024-05-08 17:21:30 -04:00

1378 lines
50 KiB
Python

from __future__ import annotations
import contextlib
import enum
import json
import logging
import uuid
from typing import (
Any,
Callable,
Dict,
Generator,
Iterable,
List,
Optional,
Tuple,
Type,
)
import numpy as np
import sqlalchemy
from langchain_core._api import deprecated, warn_deprecated
from sqlalchemy import SQLColumnExpression, delete, func
from sqlalchemy.dialects.postgresql import JSON, JSONB, UUID
from sqlalchemy.orm import Session, relationship
try:
from sqlalchemy.orm import declarative_base
except ImportError:
from sqlalchemy.ext.declarative import declarative_base
from langchain_core.documents import Document
from langchain_core.embeddings import Embeddings
from langchain_core.runnables.config import run_in_executor
from langchain_core.utils import get_from_dict_or_env
from langchain_core.vectorstores import VectorStore
from langchain_community.vectorstores.utils import maximal_marginal_relevance
class DistanceStrategy(str, enum.Enum):
"""Enumerator of the Distance strategies."""
EUCLIDEAN = "l2"
COSINE = "cosine"
MAX_INNER_PRODUCT = "inner"
DEFAULT_DISTANCE_STRATEGY = DistanceStrategy.COSINE
Base = declarative_base() # type: Any
_LANGCHAIN_DEFAULT_COLLECTION_NAME = "langchain"
class BaseModel(Base):
"""Base model for the SQL stores."""
__abstract__ = True
uuid = sqlalchemy.Column(UUID(as_uuid=True), primary_key=True, default=uuid.uuid4)
_classes: Any = None
COMPARISONS_TO_NATIVE = {
"$eq": "==",
"$ne": "!=",
"$lt": "<",
"$lte": "<=",
"$gt": ">",
"$gte": ">=",
}
SPECIAL_CASED_OPERATORS = {
"$in",
"$nin",
"$between",
}
TEXT_OPERATORS = {
"$like",
"$ilike",
}
LOGICAL_OPERATORS = {"$and", "$or"}
SUPPORTED_OPERATORS = (
set(COMPARISONS_TO_NATIVE)
.union(TEXT_OPERATORS)
.union(LOGICAL_OPERATORS)
.union(SPECIAL_CASED_OPERATORS)
)
def _get_embedding_collection_store(
vector_dimension: Optional[int] = None, *, use_jsonb: bool = True
) -> Any:
global _classes
if _classes is not None:
return _classes
from pgvector.sqlalchemy import Vector
class CollectionStore(BaseModel):
"""Collection store."""
__tablename__ = "langchain_pg_collection"
name = sqlalchemy.Column(sqlalchemy.String)
cmetadata = sqlalchemy.Column(JSON)
embeddings = relationship(
"EmbeddingStore",
back_populates="collection",
passive_deletes=True,
)
@classmethod
def get_by_name(
cls, session: Session, name: str
) -> Optional["CollectionStore"]:
return session.query(cls).filter(cls.name == name).first() # type: ignore
@classmethod
def get_or_create(
cls,
session: Session,
name: str,
cmetadata: Optional[dict] = None,
) -> Tuple["CollectionStore", bool]:
"""
Get or create a collection.
Returns [Collection, bool] where the bool is True if the collection was created.
""" # noqa: E501
created = False
collection = cls.get_by_name(session, name)
if collection:
return collection, created
collection = cls(name=name, cmetadata=cmetadata)
session.add(collection)
session.commit()
created = True
return collection, created
if use_jsonb:
# TODO(PRIOR TO LANDING): Create a gin index on the cmetadata field
class EmbeddingStore(BaseModel):
"""Embedding store."""
__tablename__ = "langchain_pg_embedding"
collection_id = sqlalchemy.Column(
UUID(as_uuid=True),
sqlalchemy.ForeignKey(
f"{CollectionStore.__tablename__}.uuid",
ondelete="CASCADE",
),
)
collection = relationship(CollectionStore, back_populates="embeddings")
embedding: Vector = sqlalchemy.Column(Vector(vector_dimension))
document = sqlalchemy.Column(sqlalchemy.String, nullable=True)
cmetadata = sqlalchemy.Column(JSONB, nullable=True)
# custom_id : any user defined id
custom_id = sqlalchemy.Column(sqlalchemy.String, nullable=True)
__table_args__ = (
sqlalchemy.Index(
"ix_cmetadata_gin",
"cmetadata",
postgresql_using="gin",
postgresql_ops={"cmetadata": "jsonb_path_ops"},
),
)
else:
# For backwards comaptibilty with older versions of pgvector
# This should be removed in the future (remove during migration)
class EmbeddingStore(BaseModel): # type: ignore[no-redef]
"""Embedding store."""
__tablename__ = "langchain_pg_embedding"
collection_id = sqlalchemy.Column(
UUID(as_uuid=True),
sqlalchemy.ForeignKey(
f"{CollectionStore.__tablename__}.uuid",
ondelete="CASCADE",
),
)
collection = relationship(CollectionStore, back_populates="embeddings")
embedding: Vector = sqlalchemy.Column(Vector(vector_dimension))
document = sqlalchemy.Column(sqlalchemy.String, nullable=True)
cmetadata = sqlalchemy.Column(JSON, nullable=True)
# custom_id : any user defined id
custom_id = sqlalchemy.Column(sqlalchemy.String, nullable=True)
_classes = (EmbeddingStore, CollectionStore)
return _classes
def _results_to_docs(docs_and_scores: Any) -> List[Document]:
"""Return docs from docs and scores."""
return [doc for doc, _ in docs_and_scores]
@deprecated(
since="0.0.31",
message=(
"This class is pending deprecation and may be removed in a future version. "
"You can swap to using the `PGVector`"
" implementation in `langchain_postgres`. "
"Please read the guidelines in the doc-string of this class "
"to follow prior to migrating as there are some differences "
"between the implementations. "
"See https://github.com/langchain-ai/langchain-postgres for details about"
"the new implementation."
),
alternative="from langchain_postgres import PGVector;",
pending=True,
)
class PGVector(VectorStore):
"""`Postgres`/`PGVector` vector store.
**DEPRECATED**: This class is pending deprecation and will likely receive
no updates. An improved version of this class is available in
`langchain_postgres` as `PGVector`. Please use that class instead.
When migrating please keep in mind that:
* The new implementation works with psycopg3, not with psycopg2
(This implementation does not work with psycopg3).
* Filtering syntax has changed to use $ prefixed operators for JSONB
metadata fields. (New implementation only uses JSONB field for metadata)
* The new implementation made some schema changes to address issues
with the existing implementation. So you will need to re-create
your tables and re-index your data or else carry out a manual
migration.
To use, you should have the ``pgvector`` python package installed.
Args:
connection_string: Postgres connection string.
embedding_function: Any embedding function implementing
`langchain.embeddings.base.Embeddings` interface.
embedding_length: The length of the embedding vector. (default: None)
NOTE: This is not mandatory. Defining it will prevent vectors of
any other size to be added to the embeddings table but, without it,
the embeddings can't be indexed.
collection_name: The name of the collection to use. (default: langchain)
NOTE: This is not the name of the table, but the name of the collection.
The tables will be created when initializing the store (if not exists)
So, make sure the user has the right permissions to create tables.
distance_strategy: The distance strategy to use. (default: COSINE)
pre_delete_collection: If True, will delete the collection if it exists.
(default: False). Useful for testing.
engine_args: SQLAlchemy's create engine arguments.
use_jsonb: Use JSONB instead of JSON for metadata. (default: True)
Strongly discouraged from using JSON as it's not as efficient
for querying.
It's provided here for backwards compatibility with older versions,
and will be removed in the future.
create_extension: If True, will create the vector extension if it doesn't exist.
disabling creation is useful when using ReadOnly Databases.
Example:
.. code-block:: python
from langchain_community.vectorstores import PGVector
from langchain_community.embeddings.openai import OpenAIEmbeddings
CONNECTION_STRING = "postgresql+psycopg2://hwc@localhost:5432/test3"
COLLECTION_NAME = "state_of_the_union_test"
embeddings = OpenAIEmbeddings()
vectorestore = PGVector.from_documents(
embedding=embeddings,
documents=docs,
collection_name=COLLECTION_NAME,
connection_string=CONNECTION_STRING,
use_jsonb=True,
)
"""
def __init__(
self,
connection_string: str,
embedding_function: Embeddings,
embedding_length: Optional[int] = None,
collection_name: str = _LANGCHAIN_DEFAULT_COLLECTION_NAME,
collection_metadata: Optional[dict] = None,
distance_strategy: DistanceStrategy = DEFAULT_DISTANCE_STRATEGY,
pre_delete_collection: bool = False,
logger: Optional[logging.Logger] = None,
relevance_score_fn: Optional[Callable[[float], float]] = None,
*,
connection: Optional[sqlalchemy.engine.Connection] = None,
engine_args: Optional[dict[str, Any]] = None,
use_jsonb: bool = False,
create_extension: bool = True,
) -> None:
"""Initialize the PGVector store."""
self.connection_string = connection_string
self.embedding_function = embedding_function
self._embedding_length = embedding_length
self.collection_name = collection_name
self.collection_metadata = collection_metadata
self._distance_strategy = distance_strategy
self.pre_delete_collection = pre_delete_collection
self.logger = logger or logging.getLogger(__name__)
self.override_relevance_score_fn = relevance_score_fn
self.engine_args = engine_args or {}
self._bind = connection if connection else self._create_engine()
self.use_jsonb = use_jsonb
self.create_extension = create_extension
if not use_jsonb:
# Replace with a deprecation warning.
warn_deprecated(
"0.0.29",
pending=True,
message=(
"Please use JSONB instead of JSON for metadata. "
"This change will allow for more efficient querying that "
"involves filtering based on metadata."
"Please note that filtering operators have been changed "
"when using JSOB metadata to be prefixed with a $ sign "
"to avoid name collisions with columns. "
"If you're using an existing database, you will need to create a"
"db migration for your metadata column to be JSONB and update your "
"queries to use the new operators. "
),
alternative=(
"Instantiate with use_jsonb=True to use JSONB instead "
"of JSON for metadata."
),
)
self.__post_init__()
def __post_init__(
self,
) -> None:
"""Initialize the store."""
if self.create_extension:
self.create_vector_extension()
EmbeddingStore, CollectionStore = _get_embedding_collection_store(
self._embedding_length, use_jsonb=self.use_jsonb
)
self.CollectionStore = CollectionStore
self.EmbeddingStore = EmbeddingStore
self.create_tables_if_not_exists()
self.create_collection()
def __del__(self) -> None:
if isinstance(self._bind, sqlalchemy.engine.Connection):
self._bind.close()
@property
def embeddings(self) -> Embeddings:
return self.embedding_function
def _create_engine(self) -> sqlalchemy.engine.Engine:
return sqlalchemy.create_engine(url=self.connection_string, **self.engine_args)
def create_vector_extension(self) -> None:
try:
with Session(self._bind) as session: # type: ignore[arg-type]
# The advisor lock fixes issue arising from concurrent
# creation of the vector extension.
# https://github.com/langchain-ai/langchain/issues/12933
# For more information see:
# https://www.postgresql.org/docs/16/explicit-locking.html#ADVISORY-LOCKS
statement = sqlalchemy.text(
"BEGIN;"
"SELECT pg_advisory_xact_lock(1573678846307946496);"
"CREATE EXTENSION IF NOT EXISTS vector;"
"COMMIT;"
)
session.execute(statement)
session.commit()
except Exception as e:
raise Exception(f"Failed to create vector extension: {e}") from e
def create_tables_if_not_exists(self) -> None:
with Session(self._bind) as session, session.begin(): # type: ignore[arg-type]
Base.metadata.create_all(session.get_bind())
def drop_tables(self) -> None:
with Session(self._bind) as session, session.begin(): # type: ignore[arg-type]
Base.metadata.drop_all(session.get_bind())
def create_collection(self) -> None:
if self.pre_delete_collection:
self.delete_collection()
with Session(self._bind) as session: # type: ignore[arg-type]
self.CollectionStore.get_or_create(
session, self.collection_name, cmetadata=self.collection_metadata
)
def delete_collection(self) -> None:
self.logger.debug("Trying to delete collection")
with Session(self._bind) as session: # type: ignore[arg-type]
collection = self.get_collection(session)
if not collection:
self.logger.warning("Collection not found")
return
session.delete(collection)
session.commit()
@contextlib.contextmanager
def _make_session(self) -> Generator[Session, None, None]:
"""Create a context manager for the session, bind to _conn string."""
yield Session(self._bind) # type: ignore[arg-type]
def delete(
self,
ids: Optional[List[str]] = None,
collection_only: bool = False,
**kwargs: Any,
) -> None:
"""Delete vectors by ids or uuids.
Args:
ids: List of ids to delete.
collection_only: Only delete ids in the collection.
"""
with Session(self._bind) as session: # type: ignore[arg-type]
if ids is not None:
self.logger.debug(
"Trying to delete vectors by ids (represented by the model "
"using the custom ids field)"
)
stmt = delete(self.EmbeddingStore)
if collection_only:
collection = self.get_collection(session)
if not collection:
self.logger.warning("Collection not found")
return
stmt = stmt.where(
self.EmbeddingStore.collection_id == collection.uuid
)
stmt = stmt.where(self.EmbeddingStore.custom_id.in_(ids))
session.execute(stmt)
session.commit()
def get_collection(self, session: Session) -> Any:
return self.CollectionStore.get_by_name(session, self.collection_name)
@classmethod
def _from(
cls,
texts: List[str],
embeddings: List[List[float]],
embedding: Embeddings,
metadatas: Optional[List[dict]] = None,
ids: Optional[List[str]] = None,
collection_name: str = _LANGCHAIN_DEFAULT_COLLECTION_NAME,
distance_strategy: DistanceStrategy = DEFAULT_DISTANCE_STRATEGY,
connection_string: Optional[str] = None,
pre_delete_collection: bool = False,
*,
use_jsonb: bool = False,
**kwargs: Any,
) -> PGVector:
if ids is None:
ids = [str(uuid.uuid4()) for _ in texts]
if not metadatas:
metadatas = [{} for _ in texts]
if connection_string is None:
connection_string = cls.get_connection_string(kwargs)
store = cls(
connection_string=connection_string,
collection_name=collection_name,
embedding_function=embedding,
distance_strategy=distance_strategy,
pre_delete_collection=pre_delete_collection,
use_jsonb=use_jsonb,
**kwargs,
)
store.add_embeddings(
texts=texts, embeddings=embeddings, metadatas=metadatas, ids=ids, **kwargs
)
return store
def add_embeddings(
self,
texts: Iterable[str],
embeddings: List[List[float]],
metadatas: Optional[List[dict]] = None,
ids: Optional[List[str]] = None,
**kwargs: Any,
) -> List[str]:
"""Add embeddings to the vectorstore.
Args:
texts: Iterable of strings to add to the vectorstore.
embeddings: List of list of embedding vectors.
metadatas: List of metadatas associated with the texts.
kwargs: vectorstore specific parameters
"""
if ids is None:
ids = [str(uuid.uuid4()) for _ in texts]
if not metadatas:
metadatas = [{} for _ in texts]
with Session(self._bind) as session: # type: ignore[arg-type]
collection = self.get_collection(session)
if not collection:
raise ValueError("Collection not found")
documents = []
for text, metadata, embedding, id in zip(texts, metadatas, embeddings, ids):
embedding_store = self.EmbeddingStore(
embedding=embedding,
document=text,
cmetadata=metadata,
custom_id=id,
collection_id=collection.uuid,
)
documents.append(embedding_store)
session.bulk_save_objects(documents)
session.commit()
return ids
def add_texts(
self,
texts: Iterable[str],
metadatas: Optional[List[dict]] = None,
ids: Optional[List[str]] = None,
**kwargs: Any,
) -> List[str]:
"""Run more texts through the embeddings and add to the vectorstore.
Args:
texts: Iterable of strings to add to the vectorstore.
metadatas: Optional list of metadatas associated with the texts.
kwargs: vectorstore specific parameters
Returns:
List of ids from adding the texts into the vectorstore.
"""
embeddings = self.embedding_function.embed_documents(list(texts))
return self.add_embeddings(
texts=texts, embeddings=embeddings, metadatas=metadatas, ids=ids, **kwargs
)
def similarity_search(
self,
query: str,
k: int = 4,
filter: Optional[dict] = None,
**kwargs: Any,
) -> List[Document]:
"""Run similarity search with PGVector with distance.
Args:
query (str): Query text to search for.
k (int): Number of results to return. Defaults to 4.
filter (Optional[Dict[str, str]]): Filter by metadata. Defaults to None.
Returns:
List of Documents most similar to the query.
"""
embedding = self.embedding_function.embed_query(text=query)
return self.similarity_search_by_vector(
embedding=embedding,
k=k,
filter=filter,
)
def similarity_search_with_score(
self,
query: str,
k: int = 4,
filter: Optional[dict] = None,
) -> List[Tuple[Document, float]]:
"""Return docs most similar to query.
Args:
query: Text to look up documents similar to.
k: Number of Documents to return. Defaults to 4.
filter (Optional[Dict[str, str]]): Filter by metadata. Defaults to None.
Returns:
List of Documents most similar to the query and score for each.
"""
embedding = self.embedding_function.embed_query(query)
docs = self.similarity_search_with_score_by_vector(
embedding=embedding, k=k, filter=filter
)
return docs
@property
def distance_strategy(self) -> Any:
if self._distance_strategy == DistanceStrategy.EUCLIDEAN:
return self.EmbeddingStore.embedding.l2_distance
elif self._distance_strategy == DistanceStrategy.COSINE:
return self.EmbeddingStore.embedding.cosine_distance
elif self._distance_strategy == DistanceStrategy.MAX_INNER_PRODUCT:
return self.EmbeddingStore.embedding.max_inner_product
else:
raise ValueError(
f"Got unexpected value for distance: {self._distance_strategy}. "
f"Should be one of {', '.join([ds.value for ds in DistanceStrategy])}."
)
def similarity_search_with_score_by_vector(
self,
embedding: List[float],
k: int = 4,
filter: Optional[dict] = None,
) -> List[Tuple[Document, float]]:
results = self._query_collection(embedding=embedding, k=k, filter=filter)
return self._results_to_docs_and_scores(results)
def _results_to_docs_and_scores(self, results: Any) -> List[Tuple[Document, float]]:
"""Return docs and scores from results."""
docs = [
(
Document(
page_content=result.EmbeddingStore.document,
metadata=result.EmbeddingStore.cmetadata,
),
result.distance if self.embedding_function is not None else None,
)
for result in results
]
return docs
def _handle_field_filter(
self,
field: str,
value: Any,
) -> SQLColumnExpression:
"""Create a filter for a specific field.
Args:
field: name of field
value: value to filter
If provided as is then this will be an equality filter
If provided as a dictionary then this will be a filter, the key
will be the operator and the value will be the value to filter by
Returns:
sqlalchemy expression
"""
if not isinstance(field, str):
raise ValueError(
f"field should be a string but got: {type(field)} with value: {field}"
)
if field.startswith("$"):
raise ValueError(
f"Invalid filter condition. Expected a field but got an operator: "
f"{field}"
)
# Allow [a-zA-Z0-9_], disallow $ for now until we support escape characters
if not field.isidentifier():
raise ValueError(
f"Invalid field name: {field}. Expected a valid identifier."
)
if isinstance(value, dict):
# This is a filter specification
if len(value) != 1:
raise ValueError(
"Invalid filter condition. Expected a value which "
"is a dictionary with a single key that corresponds to an operator "
f"but got a dictionary with {len(value)} keys. The first few "
f"keys are: {list(value.keys())[:3]}"
)
operator, filter_value = list(value.items())[0]
# Verify that that operator is an operator
if operator not in SUPPORTED_OPERATORS:
raise ValueError(
f"Invalid operator: {operator}. "
f"Expected one of {SUPPORTED_OPERATORS}"
)
else: # Then we assume an equality operator
operator = "$eq"
filter_value = value
if operator in COMPARISONS_TO_NATIVE:
# Then we implement an equality filter
# native is trusted input
native = COMPARISONS_TO_NATIVE[operator]
return func.jsonb_path_match(
self.EmbeddingStore.cmetadata,
f"$.{field} {native} $value",
json.dumps({"value": filter_value}),
)
elif operator == "$between":
# Use AND with two comparisons
low, high = filter_value
lower_bound = func.jsonb_path_match(
self.EmbeddingStore.cmetadata,
f"$.{field} >= $value",
json.dumps({"value": low}),
)
upper_bound = func.jsonb_path_match(
self.EmbeddingStore.cmetadata,
f"$.{field} <= $value",
json.dumps({"value": high}),
)
return sqlalchemy.and_(lower_bound, upper_bound)
elif operator in {"$in", "$nin", "$like", "$ilike"}:
# We'll do force coercion to text
if operator in {"$in", "$nin"}:
for val in filter_value:
if not isinstance(val, (str, int, float)):
raise NotImplementedError(
f"Unsupported type: {type(val)} for value: {val}"
)
queried_field = self.EmbeddingStore.cmetadata[field].astext
if operator in {"$in"}:
return queried_field.in_([str(val) for val in filter_value])
elif operator in {"$nin"}:
return queried_field.nin_([str(val) for val in filter_value])
elif operator in {"$like"}:
return queried_field.like(filter_value)
elif operator in {"$ilike"}:
return queried_field.ilike(filter_value)
else:
raise NotImplementedError()
else:
raise NotImplementedError()
def _create_filter_clause_deprecated(self, key, value): # type: ignore[no-untyped-def]
"""Deprecated functionality.
This is for backwards compatibility with the JSON based schema for metadata.
It uses incorrect operator syntax (operators are not prefixed with $).
This implementation is not efficient, and has bugs associated with
the way that it handles numeric filter clauses.
"""
IN, NIN, BETWEEN, GT, LT, NE = "in", "nin", "between", "gt", "lt", "ne"
EQ, LIKE, CONTAINS, OR, AND = "eq", "like", "contains", "or", "and"
value_case_insensitive = {k.lower(): v for k, v in value.items()}
if IN in map(str.lower, value):
filter_by_metadata = self.EmbeddingStore.cmetadata[key].astext.in_(
value_case_insensitive[IN]
)
elif NIN in map(str.lower, value):
filter_by_metadata = self.EmbeddingStore.cmetadata[key].astext.not_in(
value_case_insensitive[NIN]
)
elif BETWEEN in map(str.lower, value):
filter_by_metadata = self.EmbeddingStore.cmetadata[key].astext.between(
str(value_case_insensitive[BETWEEN][0]),
str(value_case_insensitive[BETWEEN][1]),
)
elif GT in map(str.lower, value):
filter_by_metadata = self.EmbeddingStore.cmetadata[key].astext > str(
value_case_insensitive[GT]
)
elif LT in map(str.lower, value):
filter_by_metadata = self.EmbeddingStore.cmetadata[key].astext < str(
value_case_insensitive[LT]
)
elif NE in map(str.lower, value):
filter_by_metadata = self.EmbeddingStore.cmetadata[key].astext != str(
value_case_insensitive[NE]
)
elif EQ in map(str.lower, value):
filter_by_metadata = self.EmbeddingStore.cmetadata[key].astext == str(
value_case_insensitive[EQ]
)
elif LIKE in map(str.lower, value):
filter_by_metadata = self.EmbeddingStore.cmetadata[key].astext.like(
value_case_insensitive[LIKE]
)
elif CONTAINS in map(str.lower, value):
filter_by_metadata = self.EmbeddingStore.cmetadata[key].astext.contains(
value_case_insensitive[CONTAINS]
)
elif OR in map(str.lower, value):
or_clauses = [
self._create_filter_clause_deprecated(key, sub_value)
for sub_value in value_case_insensitive[OR]
]
filter_by_metadata = sqlalchemy.or_(*or_clauses)
elif AND in map(str.lower, value):
and_clauses = [
self._create_filter_clause_deprecated(key, sub_value)
for sub_value in value_case_insensitive[AND]
]
filter_by_metadata = sqlalchemy.and_(*and_clauses)
else:
filter_by_metadata = None
return filter_by_metadata
def _create_filter_clause_json_deprecated(
self, filter: Any
) -> List[SQLColumnExpression]:
"""Convert filters from IR to SQL clauses.
**DEPRECATED** This functionality will be deprecated in the future.
It implements translation of filters for a schema that uses JSON
for metadata rather than the JSONB field which is more efficient
for querying.
"""
filter_clauses = []
for key, value in filter.items():
if isinstance(value, dict):
filter_by_metadata = self._create_filter_clause_deprecated(key, value)
if filter_by_metadata is not None:
filter_clauses.append(filter_by_metadata)
else:
filter_by_metadata = self.EmbeddingStore.cmetadata[key].astext == str(
value
)
filter_clauses.append(filter_by_metadata)
return filter_clauses
def _create_filter_clause(self, filters: Any) -> Any:
"""Convert LangChain IR filter representation to matching SQLAlchemy clauses.
At the top level, we still don't know if we're working with a field
or an operator for the keys. After we've determined that we can
call the appropriate logic to handle filter creation.
Args:
filters: Dictionary of filters to apply to the query.
Returns:
SQLAlchemy clause to apply to the query.
"""
if isinstance(filters, dict):
if len(filters) == 1:
# The only operators allowed at the top level are $AND and $OR
# First check if an operator or a field
key, value = list(filters.items())[0]
if key.startswith("$"):
# Then it's an operator
if key.lower() not in ["$and", "$or"]:
raise ValueError(
f"Invalid filter condition. Expected $and or $or "
f"but got: {key}"
)
else:
# Then it's a field
return self._handle_field_filter(key, filters[key])
# Here we handle the $and and $or operators
if not isinstance(value, list):
raise ValueError(
f"Expected a list, but got {type(value)} for value: {value}"
)
if key.lower() == "$and":
and_ = [self._create_filter_clause(el) for el in value]
if len(and_) > 1:
return sqlalchemy.and_(*and_)
elif len(and_) == 1:
return and_[0]
else:
raise ValueError(
"Invalid filter condition. Expected a dictionary "
"but got an empty dictionary"
)
elif key.lower() == "$or":
or_ = [self._create_filter_clause(el) for el in value]
if len(or_) > 1:
return sqlalchemy.or_(*or_)
elif len(or_) == 1:
return or_[0]
else:
raise ValueError(
"Invalid filter condition. Expected a dictionary "
"but got an empty dictionary"
)
else:
raise ValueError(
f"Invalid filter condition. Expected $and or $or "
f"but got: {key}"
)
elif len(filters) > 1:
# Then all keys have to be fields (they cannot be operators)
for key in filters.keys():
if key.startswith("$"):
raise ValueError(
f"Invalid filter condition. Expected a field but got: {key}"
)
# These should all be fields and combined using an $and operator
and_ = [self._handle_field_filter(k, v) for k, v in filters.items()]
if len(and_) > 1:
return sqlalchemy.and_(*and_)
elif len(and_) == 1:
return and_[0]
else:
raise ValueError(
"Invalid filter condition. Expected a dictionary "
"but got an empty dictionary"
)
else:
raise ValueError("Got an empty dictionary for filters.")
else:
raise ValueError(
f"Invalid type: Expected a dictionary but got type: {type(filters)}"
)
def _query_collection(
self,
embedding: List[float],
k: int = 4,
filter: Optional[Dict[str, str]] = None,
) -> List[Any]:
"""Query the collection."""
with Session(self._bind) as session: # type: ignore[arg-type]
collection = self.get_collection(session)
if not collection:
raise ValueError("Collection not found")
filter_by = [self.EmbeddingStore.collection_id == collection.uuid]
if filter:
if self.use_jsonb:
filter_clauses = self._create_filter_clause(filter)
if filter_clauses is not None:
filter_by.append(filter_clauses)
else:
# Old way of doing things
filter_clauses = self._create_filter_clause_json_deprecated(filter)
filter_by.extend(filter_clauses)
_type = self.EmbeddingStore
results: List[Any] = (
session.query(
self.EmbeddingStore,
self.distance_strategy(embedding).label("distance"), # type: ignore
)
.filter(*filter_by)
.order_by(sqlalchemy.asc("distance"))
.join(
self.CollectionStore,
self.EmbeddingStore.collection_id == self.CollectionStore.uuid,
)
.limit(k)
.all()
)
return results
def similarity_search_by_vector(
self,
embedding: List[float],
k: int = 4,
filter: Optional[dict] = None,
**kwargs: Any,
) -> List[Document]:
"""Return docs most similar to embedding vector.
Args:
embedding: Embedding to look up documents similar to.
k: Number of Documents to return. Defaults to 4.
filter (Optional[Dict[str, str]]): Filter by metadata. Defaults to None.
Returns:
List of Documents most similar to the query vector.
"""
docs_and_scores = self.similarity_search_with_score_by_vector(
embedding=embedding, k=k, filter=filter
)
return _results_to_docs(docs_and_scores)
@classmethod
def from_texts(
cls: Type[PGVector],
texts: List[str],
embedding: Embeddings,
metadatas: Optional[List[dict]] = None,
collection_name: str = _LANGCHAIN_DEFAULT_COLLECTION_NAME,
distance_strategy: DistanceStrategy = DEFAULT_DISTANCE_STRATEGY,
ids: Optional[List[str]] = None,
pre_delete_collection: bool = False,
*,
use_jsonb: bool = False,
**kwargs: Any,
) -> PGVector:
"""
Return VectorStore initialized from texts and embeddings.
Postgres connection string is required
"Either pass it as a parameter
or set the PGVECTOR_CONNECTION_STRING environment variable.
"""
embeddings = embedding.embed_documents(list(texts))
return cls._from(
texts,
embeddings,
embedding,
metadatas=metadatas,
ids=ids,
collection_name=collection_name,
distance_strategy=distance_strategy,
pre_delete_collection=pre_delete_collection,
use_jsonb=use_jsonb,
**kwargs,
)
@classmethod
def from_embeddings(
cls,
text_embeddings: List[Tuple[str, List[float]]],
embedding: Embeddings,
metadatas: Optional[List[dict]] = None,
collection_name: str = _LANGCHAIN_DEFAULT_COLLECTION_NAME,
distance_strategy: DistanceStrategy = DEFAULT_DISTANCE_STRATEGY,
ids: Optional[List[str]] = None,
pre_delete_collection: bool = False,
**kwargs: Any,
) -> PGVector:
"""Construct PGVector wrapper from raw documents and pre-
generated embeddings.
Return VectorStore initialized from documents and embeddings.
Postgres connection string is required
"Either pass it as a parameter
or set the PGVECTOR_CONNECTION_STRING environment variable.
Example:
.. code-block:: python
from langchain_community.vectorstores import PGVector
from langchain_community.embeddings import OpenAIEmbeddings
embeddings = OpenAIEmbeddings()
text_embeddings = embeddings.embed_documents(texts)
text_embedding_pairs = list(zip(texts, text_embeddings))
faiss = PGVector.from_embeddings(text_embedding_pairs, embeddings)
"""
texts = [t[0] for t in text_embeddings]
embeddings = [t[1] for t in text_embeddings]
return cls._from(
texts,
embeddings,
embedding,
metadatas=metadatas,
ids=ids,
collection_name=collection_name,
distance_strategy=distance_strategy,
pre_delete_collection=pre_delete_collection,
**kwargs,
)
@classmethod
def from_existing_index(
cls: Type[PGVector],
embedding: Embeddings,
collection_name: str = _LANGCHAIN_DEFAULT_COLLECTION_NAME,
distance_strategy: DistanceStrategy = DEFAULT_DISTANCE_STRATEGY,
pre_delete_collection: bool = False,
**kwargs: Any,
) -> PGVector:
"""
Get instance of an existing PGVector store.This method will
return the instance of the store without inserting any new
embeddings
"""
connection_string = cls.get_connection_string(kwargs)
store = cls(
connection_string=connection_string,
collection_name=collection_name,
embedding_function=embedding,
distance_strategy=distance_strategy,
pre_delete_collection=pre_delete_collection,
)
return store
@classmethod
def get_connection_string(cls, kwargs: Dict[str, Any]) -> str:
connection_string: str = get_from_dict_or_env(
data=kwargs,
key="connection_string",
env_key="PGVECTOR_CONNECTION_STRING",
)
if not connection_string:
raise ValueError(
"Postgres connection string is required"
"Either pass it as a parameter"
"or set the PGVECTOR_CONNECTION_STRING environment variable."
)
return connection_string
@classmethod
def from_documents(
cls: Type[PGVector],
documents: List[Document],
embedding: Embeddings,
collection_name: str = _LANGCHAIN_DEFAULT_COLLECTION_NAME,
distance_strategy: DistanceStrategy = DEFAULT_DISTANCE_STRATEGY,
ids: Optional[List[str]] = None,
pre_delete_collection: bool = False,
*,
use_jsonb: bool = False,
**kwargs: Any,
) -> PGVector:
"""
Return VectorStore initialized from documents and embeddings.
Postgres connection string is required
"Either pass it as a parameter
or set the PGVECTOR_CONNECTION_STRING environment variable.
"""
texts = [d.page_content for d in documents]
metadatas = [d.metadata for d in documents]
connection_string = cls.get_connection_string(kwargs)
kwargs["connection_string"] = connection_string
return cls.from_texts(
texts=texts,
pre_delete_collection=pre_delete_collection,
embedding=embedding,
distance_strategy=distance_strategy,
metadatas=metadatas,
ids=ids,
collection_name=collection_name,
use_jsonb=use_jsonb,
**kwargs,
)
@classmethod
def connection_string_from_db_params(
cls,
driver: str,
host: str,
port: int,
database: str,
user: str,
password: str,
) -> str:
"""Return connection string from database parameters."""
return f"postgresql+{driver}://{user}:{password}@{host}:{port}/{database}"
def _select_relevance_score_fn(self) -> Callable[[float], float]:
"""
The 'correct' relevance function
may differ depending on a few things, including:
- the distance / similarity metric used by the VectorStore
- the scale of your embeddings (OpenAI's are unit normed. Many others are not!)
- embedding dimensionality
- etc.
"""
if self.override_relevance_score_fn is not None:
return self.override_relevance_score_fn
# Default strategy is to rely on distance strategy provided
# in vectorstore constructor
if self._distance_strategy == DistanceStrategy.COSINE:
return self._cosine_relevance_score_fn
elif self._distance_strategy == DistanceStrategy.EUCLIDEAN:
return self._euclidean_relevance_score_fn
elif self._distance_strategy == DistanceStrategy.MAX_INNER_PRODUCT:
return self._max_inner_product_relevance_score_fn
else:
raise ValueError(
"No supported normalization function"
f" for distance_strategy of {self._distance_strategy}."
"Consider providing relevance_score_fn to PGVector constructor."
)
def max_marginal_relevance_search_with_score_by_vector(
self,
embedding: List[float],
k: int = 4,
fetch_k: int = 20,
lambda_mult: float = 0.5,
filter: Optional[Dict[str, str]] = None,
**kwargs: Any,
) -> List[Tuple[Document, float]]:
"""Return docs selected using the maximal marginal relevance with score
to embedding vector.
Maximal marginal relevance optimizes for similarity to query AND diversity
among selected documents.
Args:
embedding: Embedding to look up documents similar to.
k (int): Number of Documents to return. Defaults to 4.
fetch_k (int): Number of Documents to fetch to pass to MMR algorithm.
Defaults to 20.
lambda_mult (float): Number between 0 and 1 that determines the degree
of diversity among the results with 0 corresponding
to maximum diversity and 1 to minimum diversity.
Defaults to 0.5.
filter (Optional[Dict[str, str]]): Filter by metadata. Defaults to None.
Returns:
List[Tuple[Document, float]]: List of Documents selected by maximal marginal
relevance to the query and score for each.
"""
results = self._query_collection(embedding=embedding, k=fetch_k, filter=filter)
embedding_list = [result.EmbeddingStore.embedding for result in results]
mmr_selected = maximal_marginal_relevance(
np.array(embedding, dtype=np.float32),
embedding_list,
k=k,
lambda_mult=lambda_mult,
)
candidates = self._results_to_docs_and_scores(results)
return [r for i, r in enumerate(candidates) if i in mmr_selected]
def max_marginal_relevance_search(
self,
query: str,
k: int = 4,
fetch_k: int = 20,
lambda_mult: float = 0.5,
filter: Optional[Dict[str, str]] = None,
**kwargs: Any,
) -> List[Document]:
"""Return docs selected using the maximal marginal relevance.
Maximal marginal relevance optimizes for similarity to query AND diversity
among selected documents.
Args:
query (str): Text to look up documents similar to.
k (int): Number of Documents to return. Defaults to 4.
fetch_k (int): Number of Documents to fetch to pass to MMR algorithm.
Defaults to 20.
lambda_mult (float): Number between 0 and 1 that determines the degree
of diversity among the results with 0 corresponding
to maximum diversity and 1 to minimum diversity.
Defaults to 0.5.
filter (Optional[Dict[str, str]]): Filter by metadata. Defaults to None.
Returns:
List[Document]: List of Documents selected by maximal marginal relevance.
"""
embedding = self.embedding_function.embed_query(query)
return self.max_marginal_relevance_search_by_vector(
embedding,
k=k,
fetch_k=fetch_k,
lambda_mult=lambda_mult,
filter=filter,
**kwargs,
)
def max_marginal_relevance_search_with_score(
self,
query: str,
k: int = 4,
fetch_k: int = 20,
lambda_mult: float = 0.5,
filter: Optional[dict] = None,
**kwargs: Any,
) -> List[Tuple[Document, float]]:
"""Return docs selected using the maximal marginal relevance with score.
Maximal marginal relevance optimizes for similarity to query AND diversity
among selected documents.
Args:
query (str): Text to look up documents similar to.
k (int): Number of Documents to return. Defaults to 4.
fetch_k (int): Number of Documents to fetch to pass to MMR algorithm.
Defaults to 20.
lambda_mult (float): Number between 0 and 1 that determines the degree
of diversity among the results with 0 corresponding
to maximum diversity and 1 to minimum diversity.
Defaults to 0.5.
filter (Optional[Dict[str, str]]): Filter by metadata. Defaults to None.
Returns:
List[Tuple[Document, float]]: List of Documents selected by maximal marginal
relevance to the query and score for each.
"""
embedding = self.embedding_function.embed_query(query)
docs = self.max_marginal_relevance_search_with_score_by_vector(
embedding=embedding,
k=k,
fetch_k=fetch_k,
lambda_mult=lambda_mult,
filter=filter,
**kwargs,
)
return docs
def max_marginal_relevance_search_by_vector(
self,
embedding: List[float],
k: int = 4,
fetch_k: int = 20,
lambda_mult: float = 0.5,
filter: Optional[Dict[str, str]] = None,
**kwargs: Any,
) -> List[Document]:
"""Return docs selected using the maximal marginal relevance
to embedding vector.
Maximal marginal relevance optimizes for similarity to query AND diversity
among selected documents.
Args:
embedding (str): Text to look up documents similar to.
k (int): Number of Documents to return. Defaults to 4.
fetch_k (int): Number of Documents to fetch to pass to MMR algorithm.
Defaults to 20.
lambda_mult (float): Number between 0 and 1 that determines the degree
of diversity among the results with 0 corresponding
to maximum diversity and 1 to minimum diversity.
Defaults to 0.5.
filter (Optional[Dict[str, str]]): Filter by metadata. Defaults to None.
Returns:
List[Document]: List of Documents selected by maximal marginal relevance.
"""
docs_and_scores = self.max_marginal_relevance_search_with_score_by_vector(
embedding,
k=k,
fetch_k=fetch_k,
lambda_mult=lambda_mult,
filter=filter,
**kwargs,
)
return _results_to_docs(docs_and_scores)
async def amax_marginal_relevance_search_by_vector(
self,
embedding: List[float],
k: int = 4,
fetch_k: int = 20,
lambda_mult: float = 0.5,
filter: Optional[Dict[str, str]] = None,
**kwargs: Any,
) -> List[Document]:
"""Return docs selected using the maximal marginal relevance."""
# This is a temporary workaround to make the similarity search
# asynchronous. The proper solution is to make the similarity search
# asynchronous in the vector store implementations.
return await run_in_executor(
None,
self.max_marginal_relevance_search_by_vector,
embedding,
k=k,
fetch_k=fetch_k,
lambda_mult=lambda_mult,
filter=filter,
**kwargs,
)