changes v4

question.py

@@ -6,18 +6,17 @@ from pydantic import BaseModel
 from typing import List
 from redis_client import redis_client as r
 from dotenv import load_dotenv
-from huggingface_hub import InferenceClient
+from urllib.parse import quote_plus
 import re
+import json
 
 load_dotenv()
 
 GNEWS_API_KEY = os.getenv("GNEWS_API_KEY")
-HF_TOKEN = os.getenv("HF_TOKEN")  # Hugging Face token for private models if needed
+HF_TOKEN = os.getenv("HF_TOKEN")
 
 askMe = APIRouter()
 
-client = InferenceClient("mistralai/Mistral-7B-Instruct-v0.3", token=os.getenv("HF_TOKEN"))
-
 class QuestionInput(BaseModel):
     question: str
 
@@ -25,13 +24,16 @@ class QuestionInput(BaseModel):
 async def ask_question(input: QuestionInput):
     question = input.question
 
-    # Extract keywords (simple version)
-    keywords = re.findall(r"\b\w{4,}\b", question)
+    # Basic keyword extraction with stopword filtering
+    STOPWORDS = {"what", "about", "which", "would", "could", "this", "that"}
+    keywords = [
+        kw for kw in re.findall(r"\b\w{4,}\b", question.lower())
+        if kw not in STOPWORDS
+    ]
     query_string = " AND ".join(f'"{kw}"' for kw in keywords[:7])
+    encoded_query = quote_plus(query_string)
 
-    print("Keywords are", query_string)
-
-    gnews_url = f"https://gnews.io/api/v4/search?q={query_string}&lang=en&max=3&expand=content&token={GNEWS_API_KEY}"
+    gnews_url = f"https://gnews.io/api/v4/search?q={encoded_query}&lang=en&max=3&expand=content&token={GNEWS_API_KEY}"
     try:
         response = requests.get(gnews_url, timeout=10)
         response.raise_for_status()
@@ -39,31 +41,48 @@ async def ask_question(input: QuestionInput):
     except Exception as e:
         return {"error": f"GNews API error: {str(e)}"}
 
-    # Combine article content for context
     context = "\n\n".join([
         article.get("content") or article.get("description") or ""
         for article in articles
-    ])[:1500]  # truncate to keep model input size safe
+    ])[:1500]
 
-    print("And context is", context)
+    if not context.strip():
+        return {
+            "question": question,
+            "answer": "Cannot answer – no relevant context found.",
+            "sources": []
+        }
 
-    # Build prompt
-    prompt = f"<s>[INST] Use the context below to answer the question. If the context is insufficient, say 'Cannot answer'.\n\nContext:\n{context}\n\nQuestion: {question} [/INST]"
+    # Call HF Inference API manually
+    hf_api_url = "https://api-inference.huggingface.co/models/mistralai/Mistral-7B-Instruct-v0.3"
+    headers = {
+        "Authorization": f"Bearer {HF_TOKEN}",
+        "Content-Type": "application/json"
+    }
 
-    # result = client.text_generation(prompt, max_new_tokens=256, temperature=0.7)
+    payload = {
+        "inputs": {
+            "past_user_inputs": [],
+            "generated_responses": [],
+            "text": f"Context:\n{context}\n\nQuestion: {question}"
+        },
+        "parameters": {
+            "max_new_tokens": 256,
+            "temperature": 0.7
+        }
+    }
 
-    result = client.conversational(
-        messages=[
-            {"role": "system", "content": "You are a helpful assistant that uses provided news context to answer questions."},
-            {"role": "user", "content": f"Context:\n{context}\n\nQuestion: {question}"}
-        ],
-        max_new_tokens=256,
-        temperature=0.7
-    )
+    try:
+        response = requests.post(hf_api_url, headers=headers, data=json.dumps(payload), timeout=30)
+        response.raise_for_status()
+        hf_response = response.json()
+        answer = hf_response.get("generated_text", "Cannot answer.")
+    except Exception as e:
+        return {"error": f"Hugging Face API error: {str(e)}"}
 
     return {
         "question": question,
-        "answer": result.strip(),
+        "answer": answer.strip(),
         "sources": [
             {"title": a["title"], "url": a["url"]}
             for a in articles

requirements.txt

@@ -4,6 +4,7 @@ requests
 python-dotenv
 redis
 transformers
+accelerate
 torch
 hf_xet
 huggingface_hub