Skip to content

armgit / dspy

Public Imported from GitHub
0

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

Python

OpenAI Embeddings vectorization for KNN added

Igor Kotenkovdb6e4bcf0d77
diff --git a/dsp/modules/sentence_vectorizer.py b/dsp/modules/sentence_vectorizer.py
index 4ccc7aefc..691340843 100644
--- a/dsp/modules/sentence_vectorizer.py
+++ b/dsp/modules/sentence_vectorizer.py
@@ -1,12 +1,16 @@
import abc
-from typing import List
+from typing import List, Optional
import numpy as np
+import openai
from dsp.utils import determine_devices
class BaseSentenceVectorizer(abc.ABC):
+ # embeddings will be computed based on the string in this attribute of Example object
+ field_to_vectorize = 'text_to_vectorize'
+
def __init__(self) -> None:
pass
@@ -14,6 +18,13 @@ class BaseSentenceVectorizer(abc.ABC):
def __call__(self, inp_examples: List["Example"]) -> np.ndarray:
pass
+ def _extract_text_from_examples(self, inp_examples: List["Example"]) -> List[str]:
+ text_to_vectorize = [
+ getattr(example, self.field_to_vectorize)
+ for example in inp_examples
+ ]
+ return text_to_vectorize
+
class SentenceTransformersVectorizer(BaseSentenceVectorizer):
def __init__(
@@ -37,7 +48,8 @@ class SentenceTransformersVectorizer(BaseSentenceVectorizer):
self.normalize_embeddings = normalize_embeddings
def __call__(self, inp_examples: List["Example"]) -> np.ndarray:
- text_to_vectorize = [example.text_to_vectorize for example in inp_examples]
+ text_to_vectorize = self._extract_text_from_examples(inp_examples)
+
if self.is_gpu and self.num_devices > 1:
target_devices = list(range(self.num_devices))
pool = self.model.start_multi_process_pool(target_devices=target_devices)
@@ -63,13 +75,48 @@ class SentenceTransformersVectorizer(BaseSentenceVectorizer):
class NaiveGetFieldVectorizer(BaseSentenceVectorizer):
- def __init__(self, field_with_embeding: str = 'vectorized'):
- self.field_with_embeding = field_with_embeding
+ def __init__(self, field_with_embedding: str = 'vectorized'):
+ self.field_with_embedding = field_with_embedding
def __call__(self, inp_examples: List["Example"]) -> np.ndarray:
embeddings = [
- getattr(cur_example, self.field_with_embeding).reshape(1, -1)
+ getattr(cur_example, self.field_with_embedding).reshape(1, -1)
for cur_example in inp_examples
]
embeddings = np.concatenate(embeddings, axis=0).astype(np.float32)
return embeddings
+
+
+class OpenAIVectorizer(BaseSentenceVectorizer):
+ def __init__(
+ self, model: str = 'text-embedding-ada-002',
+ embed_batch_size: int = 1024,
+ api_key: Optional[str] = None
+ ):
+ self.model = model
+ self.embed_batch_size = embed_batch_size
+
+ if api_key:
+ openai.api_key = api_key
+
+ def __call__(self, inp_examples: List["Example"]) -> np.ndarray:
+ text_to_vectorize = self._extract_text_from_examples(inp_examples)
+ # maybe it's better to preallocate numpy matrix, but we don't know emb_dim
+ embeddings_list = []
+
+ n_batches = (len(text_to_vectorize) - 1) // self.embed_batch_size + 1
+ for cur_batch_idx in range(n_batches): # tqdm.tqdm?
+ start_idx = cur_batch_idx * self.embed_batch_size
+ end_idx = (cur_batch_idx + 1) * self.embed_batch_size
+ cur_batch = text_to_vectorize[start_idx: end_idx]
+ # cached API call:
+ response = openai.Embedding.create(
+ model=self.model,
+ input=cur_batch
+ )
+
+ cur_batch_embeddings = [cur_obj['embedding'] for cur_obj in response['data']]
+ embeddings_list.extend(cur_batch_embeddings)
+
+ embeddings = np.array(embeddings_list, dtype=np.float32)
+ return embeddings
diff --git a/knn_demonstrations_test.ipynb b/knn_demonstrations_test.ipynb
index 0920ce1a1..69e49db6a 100644
--- a/knn_demonstrations_test.ipynb
+++ b/knn_demonstrations_test.ipynb
@@ -8,6 +8,9 @@
"source": [
"import dsp\n",
"import os\n",
+ "import openai\n",
+ "# api_key = '<replace with your API key>'\n",
+ "# openai.api_key = api_key\n",
"\n",
"os.environ['KMP_DUPLICATE_LIB_OK'] = 'TRUE'"
]
@@ -179,7 +182,6 @@
"# context manager is working properly\n",
"with dsp.settings.context(vectorizer=dsp.SentenceTransformersVectorizer()):\n",
" knn_func = dsp.knn(train)\n",
- " \n",
"\n",
" knn_res_train_vec = knn_func(train[QUERY_IDX], 3)\n",
" print(train[QUERY_IDX].question) # the query itself\n",
@@ -197,7 +199,7 @@
{
"data": {
"text/plain": [
- "<dsp.modules.sentence_vectorizer.NaiveGetFieldVectorizer at 0x7fc21559c130>"
+ "<dsp.modules.sentence_vectorizer.NaiveGetFieldVectorizer at 0x7faae3d71fd0>"
]
},
"execution_count": 11,
@@ -249,7 +251,43 @@
},
{
"cell_type": "code",
- "execution_count": null,
+ "execution_count": 14,
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "<dsp.modules.sentence_vectorizer.OpenAIVectorizer object at 0x7faac15c6190>\n",
+ "Who produced the album that included a re-recording of \"Lithium\"?\n",
+ "--------------------------------------------------------------------------------\n",
+ "Who produced the album that included a re-recording of \"Lithium\"?\n",
+ "There were two founders of recording studio called \"Smart Studios\". The first one was Butch Vig. Who was the second founder?\n",
+ "Who founded Smart Studios in Madison, while still performing drums in Spooner at night and driving a taxi cab during the day?\n"
+ ]
+ }
+ ],
+ "source": [
+ "QUERY_IDX = 0\n",
+ "\n",
+ "# let's try OpenAI embeddings\n",
+ "with dsp.settings.context(vectorizer=dsp.OpenAIVectorizer()):\n",
+ " print(dsp.settings.vectorizer)\n",
+ " knn_func = dsp.knn(train)\n",
+ "\n",
+ " knn_res_train_vec = knn_func(train[QUERY_IDX], 3)\n",
+ " print(train[QUERY_IDX].question) # the query itself\n",
+ " # similar questions\n",
+ " print('-'*80)\n",
+ " for i in [train[train.index(i)].question for i in knn_res_train_vec]:\n",
+ " print(i)\n",
+ "\n",
+ "# the order is different from embeddings from SentenceTransformers"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 2,
"metadata": {},
"outputs": [],
"source": []