Skip to content

armgit / dspy

Public Imported from GitHub
0

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

Python

Fix pgvector work with the tutorial

Hieu Nguyen Van0b5ff452f33e
diff --git a/dspy/retrieve/pgvector_rm.py b/dspy/retrieve/pgvector_rm.py
index 72069ec86..ae8d54f1a 100644
--- a/dspy/retrieve/pgvector_rm.py
+++ b/dspy/retrieve/pgvector_rm.py
@@ -14,8 +14,10 @@ except ImportError:
try:
import openai
except ImportError:
- warnings.warn("`openai` is not installed. Install it with `pip install openai` to use OpenAI embedding models.",
- category=ImportWarning)
+ warnings.warn(
+ "`openai` is not installed. Install it with `pip install openai` to use OpenAI embedding models.",
+ category=ImportWarning,
+ )
class PgVectorRM(dspy.Retrieve):
@@ -32,6 +34,7 @@ class PgVectorRM(dspy.Retrieve):
pg_table_name (Optional[str]): name of the table containing passages
openai_client (openai.OpenAI): OpenAI client to use for computing query embeddings. Either openai_client or embedding_func must be provided.
embedding_func (Callable): A function to use for computing query embeddings. Either openai_client or embedding_func must be provided.
+ content_field (str = "text"): Field containing the passage text. Defaults to "text"
k (Optional[int]): Default number of top passages to retrieve. Defaults to 20
embedding_field (str = "embedding"): Field containing passage embeddings. Defaults to "embedding"
fields (List[str] = ['text']): Fields to retrieve from the table. Defaults to "text"
@@ -62,41 +65,46 @@ class PgVectorRM(dspy.Retrieve):
self.retrieve = PgVectorRM(db_url, openai_client=openai_client, "paragraphs", fields=["text", "document_id"], k=20)
```
"""
+
def __init__(
- self,
- db_url: str,
- pg_table_name: str,
- openai_client: Optional[openai.OpenAI] = None,
- embedding_func: Optional[Callable] = None,
- k: int = 20,
- embedding_field: str = "embedding",
- fields: Optional[list[str]] = None,
- embedding_model: str = "text-embedding-ada-002",
- include_similarity: bool = False,
+ self,
+ db_url: str,
+ pg_table_name: str,
+ openai_client: Optional[openai.OpenAI] = None,
+ embedding_func: Optional[Callable] = None,
+ k: int = 20,
+ embedding_field: str = "embedding",
+ fields: Optional[list[str]] = None,
+ content_filed: str = "text",
+ embedding_model: str = "text-embedding-ada-002",
+ include_similarity: bool = False,
):
"""
k = 20 is the number of paragraphs to retrieve
"""
- assert openai_client or embedding_func, "Either openai_client or embedding_func must be provided."
+ assert (
+ openai_client or embedding_func
+ ), "Either openai_client or embedding_func must be provided."
self.openai_client = openai_client
self.embedding_func = embedding_func
self.conn = psycopg2.connect(db_url)
register_vector(self.conn)
self.pg_table_name = pg_table_name
- self.fields = fields or ['text']
+ self.fields = fields or ["text"]
+ self.content_filed = content_filed
self.embedding_field = embedding_field
self.embedding_model = embedding_model
self.include_similarity = include_similarity
super().__init__(k=k)
- def forward(self, query: str):
- """Search with PgVector for self.k top passages for query using cosine similarity
+ def forward(self, query: str, k: int = None):
+ """Search with PgVector for k top passages for query using cosine similarity
Args:
query (str): The query to search for
- include_similarity (bool): Whether or not to include the similarity for each record
+ k (int): The number of top passages to retrieve. Defaults to the value set in the constructor.
Returns:
dspy.Prediction: an object containing the retrieved passages.
"""
@@ -105,21 +113,15 @@ class PgVectorRM(dspy.Retrieve):
retrieved_docs = []
- fields = sql.SQL(',').join([
- sql.Identifier(f)
- for f in self.fields
- ])
+ fields = sql.SQL(",").join([sql.Identifier(f) for f in self.fields])
if self.include_similarity:
- similarity_field = (
- sql.SQL(',') +
- sql.SQL(
- '1 - ({embedding_field} <=> %s) AS similarity',
- ).format(embedding_field=sql.Identifier(self.embedding_field))
- )
+ similarity_field = sql.SQL(",") + sql.SQL(
+ "1 - ({embedding_field} <=> %s::vector) AS similarity",
+ ).format(embedding_field=sql.Identifier(self.embedding_field))
fields += similarity_field
- args = (query_embedding, query_embedding, self.k)
+ args = (query_embedding, query_embedding, k if k else self.k)
else:
- args = (query_embedding, self.k)
+ args = (query_embedding, k if k else self.k)
sql_query = sql.SQL(
"select {fields} from {table} order by {embedding_field} <=> %s::vector limit %s",
@@ -131,23 +133,26 @@ class PgVectorRM(dspy.Retrieve):
with self.conn as conn:
with conn.cursor() as cur:
- cur.execute(
- sql_query,
- args)
+ cur.execute(sql_query, args)
rows = cur.fetchall()
columns = [descrip[0] for descrip in cur.description]
for row in rows:
data = dict(zip(columns, row))
+ data["long_text"] = data[self.content_filed]
retrieved_docs.append(dspy.Example(**data))
# Return Prediction
return retrieved_docs
def _get_embeddings(self, query: str) -> list[float]:
if self.openai_client is not None:
- return self.openai_client.embeddings.create(
- model=self.embedding_model,
- input=query,
- encoding_format="float",
- ).data[0].embedding
+ return (
+ self.openai_client.embeddings.create(
+ model=self.embedding_model,
+ input=query,
+ encoding_format="float",
+ )
+ .data[0]
+ .embedding
+ )
else:
return self.embedding_func(query)