feed zeros for required token_type_ids input

Dockerfile

@@ -24,7 +24,8 @@ tok=AutoTokenizer.from_pretrained(m); \
 sess=ort.InferenceSession(hf_hub_download(m,'model.onnx'),providers=['CPUExecutionProvider']); \
 names={i.name for i in sess.get_inputs()}; \
 e=tok(['ભાવ સમાચાર','mandi prices'],padding=True,truncation=True,max_length=512,return_tensors='np'); \
-h=sess.run(None,{k:v for k,v in e.items() if k in names})[0]; \
+f={n:(e[n] if n in e else np.zeros_like(e['input_ids'])) for n in names}; \
+h=sess.run(None,f)[0]; \
 c=h[:,0]; c=c/np.linalg.norm(c,axis=1,keepdims=True); \
 assert c.shape==(2,1024), c.shape; \
 assert abs(float(np.linalg.norm(c[0]))-1.0)<1e-3; \

app.py

@@ -57,8 +57,11 @@ def _encode(texts: list[str]) -> list[list[float]]:
             chunk, padding=True, truncation=True,
             max_length=MAX_TOKENS, return_tensors="np",
         )
-        # bge-m3 (xlm-roberta) has no token_type_ids; feed only what the graph wants.
-        feed = {k: v for k, v in enc.items() if k in _input_names}
+        # The ONNX graph requires token_type_ids but the xlm-roberta tokenizer
+        # doesn't emit them (bge-m3 ignores them) → feed zeros for any required
+        # input the tokenizer didn't produce.
+        ids = enc["input_ids"]
+        feed = {n: enc[n] if n in enc else np.zeros_like(ids) for n in _input_names}
         hidden = _session.run(None, feed)[0]  # (B, T, 1024) last_hidden_state
         # dense embedding = CLS token (position 0), then L2-normalize so cosine == dot.
         cls = hidden[:, 0]