mr / app /engine /chunk_embed.py
JPBianchi's picture
endpoint only, no UI
ae92cb7
raw
history blame
2.26 kB
import os
import pandas as pd
import torch
from app.settings import parquet_file
import tiktoken # tokenizer library for use with OpenAI LLMs
from llama_index.legacy.text_splitter import SentenceSplitter
from sentence_transformers import SentenceTransformer
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# create tensors on GPU if available
if torch.cuda.is_available():
torch.set_default_tensor_type('torch.cuda.FloatTensor')
def chunk_vectorize(doc_content: dict = None,
chunk_size: int = 256, # limit for 'all-mpnet-base-v2'
chunk_overlap: int = 20, # some overlap to link the chunks
encoder: str = 'gpt-3.5-turbo-0613',
model_name: str = 'sentence-transformers/all-mpnet-base-v2'): # can try all-MiniLM-L6-v2
# see tests in chunking_indexing.ipynb for more details
encoding = tiktoken.encoding_for_model(encoder)
splitter = SentenceSplitter(chunk_size=chunk_size,
tokenizer=encoding.encode,
chunk_overlap=chunk_overlap)
# let's create the splits for every document
contents_splits = {}
for fname, content in doc_content.items():
splits = [splitter.split_text(page) for page in content]
contents_splits[fname] = [split for sublist in splits for split in sublist]
model = SentenceTransformer(model_name)
content_emb = {}
for fname, splits in contents_splits.items():
content_emb[fname] = [(split, model.encode(split)) for split in splits]
# save fname since it carries information, and could be used as a property in Weaviate
text_vector_tuples = [(fname, split, emb.tolist()) for fname, splits_emb in content_emb.items() for split, emb in splits_emb]
new_df = pd.DataFrame(
text_vector_tuples,
columns=['file', 'content', 'content_embedding']
)
# load the existing parquet file if it exists and update it
if os.path.exists(parquet_file):
new_df = pd.concat([pd.read_parquet(parquet_file), new_df])
# no optimization here (zipping etc) since the data is small
new_df.to_parquet(parquet_file, index=False)
return