Fix Vectara pipeline: explicitly load tokenizer before pipeline init

transformers<4.46 can't auto-detect the tokenizer for Vectara's custom
HHEMv2 architecture, leaving pipeline.tokenizer as None at inference.
Loading AutoTokenizer explicitly and passing it in resolves the TypeError.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Dockerfile

@@ -13,9 +13,10 @@ RUN pip install --no-cache-dir -r requirements.txt
 # Pre-download models so first request isn't slow on HF Spaces
 RUN python -c "\
 from sentence_transformers import SentenceTransformer; \
-from transformers import pipeline; \
+from transformers import AutoTokenizer, pipeline; \
 SentenceTransformer('all-MiniLM-L6-v2'); \
-pipeline('text-classification', model='vectara/hallucination_evaluation_model', trust_remote_code=True)"
+tok = AutoTokenizer.from_pretrained('vectara/hallucination_evaluation_model', trust_remote_code=True); \
+pipeline('text-classification', model='vectara/hallucination_evaluation_model', tokenizer=tok, trust_remote_code=True)"
 
 COPY knowledge/ ./knowledge/
 COPY backend/   ./backend/

backend/grader.py

@@ -16,7 +16,7 @@ from typing import Any
 
 from sentence_transformers import SentenceTransformer
 from sklearn.metrics.pairwise import cosine_similarity
-from transformers import pipeline as hf_pipeline
+from transformers import AutoTokenizer, pipeline as hf_pipeline
 
 from config import EMBEDDER_MODEL
 from rosetta import check_terminology
@@ -41,9 +41,11 @@ def get_nli_model() -> Any:
     """Return the shared Vectara faithfulness pipeline, loading it on first call."""
     global _nli_model
     if _nli_model is None:
+        tokenizer = AutoTokenizer.from_pretrained(NLI_MODEL, trust_remote_code=True)
         _nli_model = hf_pipeline(
             "text-classification",
             model=NLI_MODEL,
+            tokenizer=tokenizer,
             trust_remote_code=True,
             truncation=True,
             max_length=512,