Skip to content

armgit / dspy

Public Imported from GitHub
0

DSPy: The framework for programming—not prompting—language models

Python

Support returning similarity

Josh Mantovani86a869be9d04
diff --git a/dspy/retrieve/pgvector_rm.py b/dspy/retrieve/pgvector_rm.py
index 942d1df70..ccce60523 100644
--- a/dspy/retrieve/pgvector_rm.py
+++ b/dspy/retrieve/pgvector_rm.py
@@ -89,12 +89,12 @@ class PgVectorRM(dspy.Retrieve):
super().__init__(k=k)
- def forward(self, query: str):
- """Search with PgVector for self.k top passages for query
+ def forward(self, query: str, include_similarity: bool = False):
+ """Search with PgVector for self.k top passages for query using cosine similarity
Args:
query (str): The query to search for
- k (Optional[int]): The number of top passages to retrieve. Defaults to self.k
+ include_similarity (bool): Whether or not to include the similarity for each record
Returns:
dspy.Prediction: an object containing the retrieved passages.
"""
@@ -103,12 +103,23 @@ class PgVectorRM(dspy.Retrieve):
related_paragraphs = []
+ fields = sql.SQL(',').join([
+ sql.Identifier(f)
+ for f in self.fields
+ ])
+ if include_similarity:
+ similarity_field = (
+ sql.SQL(',') +
+ sql.SQL(
+ '1 - ({embedding_field} <=> %s) AS similarity'
+ ).format(embedding_field=sql.Identifier(self.embedding_field))
+ )
+ fields += similarity_field
+
sql_query = sql.SQL(
- "select {fields} from {table} order by {embedding_field} <-> %s::vector limit %s").format(
- fields=sql.SQL(',').join([
- sql.Identifier(f)
- for f in self.fields
- ]),
+ "select {fields} from {table} order by {embedding_field} <=> %s::vector limit %s"
+ ).format(
+ fields=fields,
table=sql.Identifier(self.pg_table_name),
embedding_field=sql.Identifier(self.embedding_field),
)
@@ -117,7 +128,7 @@ class PgVectorRM(dspy.Retrieve):
with conn.cursor() as cur:
cur.execute(
sql_query,
- (query_embedding, self.k))
+ (query_embedding, query_embedding, self.k))
rows = cur.fetchall()
columns = [descrip[0] for descrip in cur.description]
for row in rows: