Initial codes commit

models/embedder.py

@@ -3,6 +3,7 @@ from typing import List
 import numpy as np
 import onnxruntime as ort
 from fastapi import Request
+from modules.utils import generate_position_ids
 
 def _l2_normalize(vec: np.ndarray) -> List[float]:
     norm = np.linalg.norm(vec) or 1.0
@@ -17,6 +18,8 @@ def get_embedding(request: Request, text: str) -> List[float]:
     sess: ort.InferenceSession = request.app.state.embedder_sess
 
     inputs = tokenizer(text, return_tensors="np", padding=True, truncation=True, max_length=256)
+    input_ids = inputs["input_ids"]
+    inputs["position_ids"] = generate_position_ids(input_ids)
     ort_inputs = {k: v for k, v in inputs.items()}
     ort_outs = sess.run(None, ort_inputs)
     # 일반적으로 첫 번째 출력이 [batch, dim] 임베딩

models/reranker.py

@@ -3,6 +3,7 @@ import os
 from typing import List, Dict
 import onnxruntime as ort
 from fastapi import Request
+from modules.utils import generate_position_ids
 
 THRESHOLD = float(os.getenv("RERANK_THRESHOLD", "0.3"))
 
@@ -20,6 +21,8 @@ def rerank(request: Request, query: str, contexts: List[Dict]) -> List[Dict]:
 
     pairs = [(query, ctx["text"]) for ctx in contexts]
     inputs = tokenizer(pairs, return_tensors="np", padding=True, truncation=True, max_length=256)
+    input_ids = inputs["input_ids"]
+    inputs["position_ids"] = generate_position_ids(input_ids)
     ort_inputs = {k: v for k, v in inputs.items()}
     scores = sess.run(None, ort_inputs)[0]  # [batch, 1] 형태
     scores = scores.squeeze(-1)

modules/utils.py

@@ -1,5 +1,6 @@
 # rag/modules/utils.py
 import os
+import numpy as np
 
 def ensure_dir(path: str):
     os.makedirs(path, exist_ok=True)
@@ -10,3 +11,12 @@ def touch(path: str):
 
 def exists(path: str) -> bool:
     return os.path.exists(path)
+
+def generate_position_ids(input_ids: np.ndarray) -> np.ndarray:
+    """
+    input_ids: [batch_size, seq_len]
+    return: position_ids of shape [batch_size, seq_len] with int64 dtype
+    """
+    batch_size, seq_len = input_ids.shape
+    position_ids = np.arange(seq_len)[None, :].astype("int64")
+    return np.tile(position_ids, (batch_size, 1))