The embedding is called twice

[Chen-Zile12]

AutoSchemaKG\atlas_rag\vectorstore\create_graph_index.py

def compute_text_embeddings(text_list, sentence_encoder: BaseEmbeddingModel, batch_size = 40, normalize_embeddings: bool = False):
"""Separated text embedding computation"""
text_embeddings = []

for i in tqdm(range(0, len(text_list), batch_size), desc="Encoding texts"):
    batch = text_list[i:i + batch_size]
    **embeddings = sentence_encoder.encode(batch, normalize_embeddings=normalize_embeddings)**
    if isinstance(embeddings, torch.Tensor):
        embeddings = embeddings.cpu().numpy()
    **text_embeddings.extend(sentence_encoder.encode(batch, normalize_embeddings = normalize_embeddings))**
return text_embeddings

[gzone0111]

Good catch, thank you for spotting this!

The bug is in compute_text_embeddings() at create_graph_index.py:47. After computing embeddings = sentence_encoder.encode(batch, ...) on line 44, the result is stored but never used — line 47 calls sentence_encoder.encode() again on the same batch:

embeddings = sentence_encoder.encode(batch, ...)          # result unused
if isinstance(embeddings, torch.Tensor):
    embeddings = embeddings.cpu().numpy()                  # wasted conversion
text_embeddings.extend(sentence_encoder.encode(batch, ...)) # second call!

The sibling function compute_graph_embeddings() does not have this bug (line 15 correctly uses the single call result directly).

Fix: Changed line 47 to text_embeddings.extend(embeddings), reusing the already-computed result. This halves the embedding time for text lists.