Skip to content

armgit / dspy

Public Imported from GitHub
0

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

Python

Docstrings & small fixes

Igor Kotenkovabaf8d7e5b7f
diff --git a/dsp/modules/sentence_vectorizer.py b/dsp/modules/sentence_vectorizer.py
index 691340843..f45d3b696 100644
--- a/dsp/modules/sentence_vectorizer.py
+++ b/dsp/modules/sentence_vectorizer.py
@@ -8,6 +8,12 @@ from dsp.utils import determine_devices
class BaseSentenceVectorizer(abc.ABC):
+ '''
+ Base Class for Vectorizers. The main purpose is to vectorize text (doc/query)
+ for ANN/KNN indexes. `__call__` method takes `List[Example]` as a single input, then extracts
+ `field_to_vectorize` from every Example and convert them into embeddings.
+ You can customize extraction logic in the `_extract_text_from_examples` method.
+ '''
# embeddings will be computed based on the string in this attribute of Example object
field_to_vectorize = 'text_to_vectorize'
@@ -27,6 +33,12 @@ class BaseSentenceVectorizer(abc.ABC):
class SentenceTransformersVectorizer(BaseSentenceVectorizer):
+ '''
+ Vectorizer bsaed on `SentenceTransformers` models. You can pick any model from this link:
+ https://huggingface.co/sentence-transformers
+ More details about models:
+ https://www.sbert.net/docs/pretrained_models.html
+ '''
def __init__(
self,
model_name_or_path: str = 'all-MiniLM-L6-v2',
@@ -75,6 +87,10 @@ class SentenceTransformersVectorizer(BaseSentenceVectorizer):
class NaiveGetFieldVectorizer(BaseSentenceVectorizer):
+ '''
+ If embeddings were precomputed, then we could just extract them from the proper field
+ (set by `field_with_embedding`) from each `Example`.
+ '''
def __init__(self, field_with_embedding: str = 'vectorized'):
self.field_with_embedding = field_with_embedding
@@ -88,8 +104,14 @@ class NaiveGetFieldVectorizer(BaseSentenceVectorizer):
class OpenAIVectorizer(BaseSentenceVectorizer):
+ '''
+ This vectorizer uses OpenAI API to convert texts to embeddings. Changing `model` is not
+ recommended. More about the model: https://openai.com/blog/new-and-improved-embedding-model/
+ `api_key` should be passed as an argument or as env variable (`OPENAI_API_KEY`).
+ '''
def __init__(
- self, model: str = 'text-embedding-ada-002',
+ self,
+ model: str = 'text-embedding-ada-002',
embed_batch_size: int = 1024,
api_key: Optional[str] = None
):
@@ -109,7 +131,7 @@ class OpenAIVectorizer(BaseSentenceVectorizer):
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:
+ # OpenAI API call:
response = openai.Embedding.create(
model=self.model,
input=cur_batch
diff --git a/dsp/primitives/demonstrate.py b/dsp/primitives/demonstrate.py
index 7d42003eb..10a725a47 100644
--- a/dsp/primitives/demonstrate.py
+++ b/dsp/primitives/demonstrate.py
@@ -75,8 +75,11 @@ def annotate(*transformations):
def sample(train, k: int):
- branch_idx = hasattr(dsp.settings, 'branch_idx') and isinstance(
- dsp.settings.branch_idx, int) and dsp.settings.branch_idx
+ branch_idx = (
+ hasattr(dsp.settings, 'branch_idx')
+ and isinstance(dsp.settings.branch_idx, int)
+ and dsp.settings.branch_idx
+ )
branch_idx = branch_idx or 0
# print(f"branch_idx = {branch_idx}")
@@ -93,8 +96,13 @@ def sample(train, k: int):
def all_but(train, x):
# output = [y for y in train if y.question != x.question]
- output = [y for y in train
- if not set.intersection(set(x.get('history', []) + [x.question]), set(y.get('history', []) + [y.question]))]
+ output = [
+ y for y in train
+ if not set.intersection(
+ set(x.get('history', []) + [x.question]),
+ set(y.get('history', []) + [y.question])
+ )
+ ]
# print(len(output))
return output
@@ -113,14 +121,25 @@ def answer_match(prediction, answers):
def passage_has_answers(passage, answers):
- return has_answer([DPR_normalize(normalize_text(ans)) for ans in answers], normalize_text(passage))
+ return has_answer(
+ tokenized_answers=[DPR_normalize(normalize_text(ans)) for ans in answers],
+ text=normalize_text(passage)
+ )
def cast_naive_get_only_question_text(inp_example: Example) -> Example:
+ '''
+ Extracts question as a field to vectorize with Vectorizer object. `question` field is used.
+ '''
return inp_example.copy(text_to_vectorize=inp_example.question)
def cast_naive_get_question_and_answer(inp_example: Example) -> Example:
+ '''
+ Extracts question and answer as fields to vectorize with Vectorizer object.
+ `question` and `answer` fields are used. They will be concatenated with the word "Answer"
+ between.
+ '''
text_to_vectorize = inp_example.question.strip() + " Answer: " + inp_example.answer.strip()
return inp_example.copy(text_to_vectorize=text_to_vectorize)
@@ -130,6 +149,19 @@ def knn(
cast: Callable[[Example], Example] = cast_naive_get_only_question_text,
**knn_args
) -> Callable[[Example, int], List[Example]]:
+ '''
+ A function that vectorizes train data using `dsm.settings.vectorizer`, then build an ANN/KNN
+ index to search similar questions among `train` samples.
+
+ Args:
+ train: a bunch of questions to put in index & search later
+ cast: function that contructs text before vectorization. By default,
+ it uses only question. Check `cast_naive_get_question_and_answer` for more details.
+ n_probe: number of closest IVF-clusters to check for neighbours.
+ Doesn't affect bruteforce-based search.
+ knn_args: check `create_faiss_index` function for details on ANN/KNN arguments.
+ Returns: function to search similar Examples from `train` in FAISS-index.
+ '''
train_casted_to_vectorize = [cast(cur_elem) for cur_elem in train]
vectorizer: "BaseSentenceVectorizer" = dsp.settings.vectorizer