Flora Aligner Bugs

[abhishekreddyallu]

FLORA aligner workflows have default model and cached-embedding inconsistencies

1. Default FLORAAligner passes None as the embedding model ID

Location: ontoaligner/aligner/flora/flora.py:192, ontoaligner/aligner/flora/flora.py:241-246

Current code:

model_id: Optional[str] = None,

and:

self.literals_embedding = FLORALiteralsEmbedding(
    model_id=model_id,
    device=device if device else 'cpu',
    identity=string_identity,
    emb_path=emb_path,
)

This explicitly forwards model_id=None to FLORALiteralsEmbedding, which overrides the embedding class default model:

model_id: str = "Lihuchen/pearl_small"

Observed before fix:

Creating FLORAAligner() with public defaults...
FAIL: FLORAAligner() failed with public defaults.
Error type: OSError
Error message: None is not a local folder and is not a valid model identifier listed on 'https://huggingface.co/models'
If this is a private repository, make sure to pass a token having permission to this repo either by logging in with `hf auth login` or by passing `token=<your_token>`

Fix:

Use the same concrete default in FLORAAligner:

model_id: Optional[str] = "Lihuchen/pearl_small",

This keeps default construction valid while still allowing users to pass their own model:

FLORAAligner(model_id="custom-model-id")

Verified after fix:

PASS: FLORAAligner() constructed successfully.
Aligner: FLORAAligner

---

2. Cached embeddings are assigned as tuples to both graph slots

Location: ontoaligner/aligner/flora/literals.py:407-410, ontoaligner/aligner/flora/literals.py:633-646

Current code:

self.kg1_pretrained = self.load_embeddings(emb_path)
self.kg2_pretrained = self.load_embeddings(emb_path)

But load_embeddings() returns two values:

def load_embeddings(
    self,
    emb_path: str,
) -> Tuple[Dict[str, Any], Dict[str, Any]]:
    """Load precomputed embeddings for two knowledge bases.

    Args:
        emb_path: Directory path where embedding files are saved.
    """
    with open(os.path.join(emb_path, "kb1.pkl"), "rb") as f:
        kb1_emb = pickle.load(f)
    with open(os.path.join(emb_path, "kb2.pkl"), "rb") as f:
        kb2_emb = pickle.load(f)
    return kb1_emb, kb2_emb

So the current code assigns the full tuple to both attributes:

self.kg1_pretrained = (kb1_emb, kb2_emb)
self.kg2_pretrained = (kb1_emb, kb2_emb)

This breaks the documented cached-embedding workflow because each graph slot should receive its own embedding dictionary.

Fix:

Unpack the returned tuple once:

self.kg1_pretrained, self.kg2_pretrained = self.load_embeddings(emb_path)

This assigns:

self.kg1_pretrained = kb1_emb
self.kg2_pretrained = kb2_emb

instead of assigning the full (kb1_emb, kb2_emb) tuple to both.