Upload 34 files

DEPLOY.md

@@ -74,6 +74,16 @@ real LLM responses:
 4. Restart the Space (**Settings → Factory reboot**, or just wait — it
    picks up new secrets on the next restart).
 
+**Note on the API endpoint:** `src/llm_client.py` explicitly passes
+`provider="hf-inference"` to `InferenceClient`, and `requirements.txt`
+pins a recent `huggingface_hub` version. Both matter: older client
+versions / omitting the provider can silently route calls through the
+now-deprecated `api-inference.huggingface.co` domain, which fails with a
+DNS resolution error rather than a clear auth error. If you ever see a
+`NameResolutionError` mentioning `api-inference.huggingface.co` in the
+"Test LLM connection" diagnostics, bump `huggingface_hub` in
+`requirements.txt` to the latest release.
+
 ## Re-training / updating the models
 
 The Space **loads pre-built artifacts** from `models/` and `data/` — it

requirements.txt

@@ -4,5 +4,5 @@ pandas==2.2.3
 numpy==1.26.4
 matplotlib==3.9.2
 joblib==1.4.2
-huggingface_hub==0.26.5
+huggingface_hub==0.31.4
 spaces>=0.30.0

src/llm_client.py

@@ -125,7 +125,12 @@ def answer_query(
     errors = []
     for candidate in MODEL_CANDIDATES:
         try:
-            client = InferenceClient(model=candidate, token=hf_token)
+            # `provider="hf-inference"` forces routing through HF's current
+            # Inference Providers router (router.huggingface.co). Without it,
+            # older huggingface_hub versions / ambiguous setups can fall back
+            # to the deprecated `api-inference.huggingface.co` endpoint,
+            # which has been sunset and fails with a DNS resolution error.
+            client = InferenceClient(model=candidate, token=hf_token, provider="hf-inference")
             completion = client.chat_completion(messages=messages, max_tokens=max_tokens, temperature=0.3)
             text = completion.choices[0].message.content
             if text and text.strip():