Create app.py

app.py

@@ -0,0 +1,79 @@
+import os
+os.environ["TRANSFORMERS_CACHE"] = "/tmp"
+os.environ["HF_HOME"] = "/tmp"
+
+from fastapi import FastAPI, HTTPException
+from pydantic import BaseModel
+import json
+import torch
+from transformers import MT5ForConditionalGeneration, MT5Tokenizer
+from sentence_transformers import SentenceTransformer, util
+
+# Load dataset
+with open("data/gpt2_ready_filtered.jsonl", "r", encoding="utf-8") as f:
+    data = [json.loads(line) for line in f]
+texts = [item["text"] for item in data]
+
+# Load model
+model_name = "nurfarah57/SQ-MT5"
+tokenizer = MT5Tokenizer.from_pretrained(model_name)
+model = MT5ForConditionalGeneration.from_pretrained(model_name)
+model.eval()
+
+# Load sentence embedder
+embedder = SentenceTransformer("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2")
+embeddings = embedder.encode(texts, convert_to_tensor=True)
+
+# FastAPI app
+app = FastAPI(
+    title="Somali QA API",
+    description="Su’aal weydii oo hel jawaab laga raadshay dataset-ka ama laga sameeyay model.",
+    version="1.0"
+)
+
+# Input schema
+class QuestionRequest(BaseModel):
+    question: str
+
+# Extract question/answer from dataset line
+def extract_qa(text):
+    parts = text.split("\nJawaab:")
+    if len(parts) == 2:
+        return parts[0].replace("Su'aal:", "").strip(), parts[1].strip()
+    return None, None
+
+# Match dataset semantically
+def find_semantic_match(question, threshold=0.90):
+    user_emb = embedder.encode(question, convert_to_tensor=True)
+    hits = util.semantic_search(user_emb, embeddings, top_k=1)
+    if hits and hits[0][0]["score"] >= threshold:
+        idx = hits[0][0]["corpus_id"]
+        _, jawaab = extract_qa(texts[idx])
+        return jawaab
+    return None
+
+# Fallback generation
+def generate_with_mt5(question):
+    prompt = f"su'aal: {question}"
+    inputs = tokenizer(prompt, return_tensors="pt", truncation=True)
+    with torch.no_grad():
+        outputs = model.generate(inputs["input_ids"], max_length=128)
+    return tokenizer.decode(outputs[0], skip_special_tokens=True)
+
+# API endpoint
+@app.post("/qa")
+def answer_question(req: QuestionRequest):
+    if not req.question.strip():
+        raise HTTPException(status_code=400, detail="Su’aal lama helin.")
+
+    match = find_semantic_match(req.question)
+    if match:
+        return {"answer": match, "source": "dataset"}
+
+    generated = generate_with_mt5(req.question)
+    return {"answer": generated, "source": "model"}
+
+# Root
+@app.get("/")
+def root():
+    return {"message": "✅ Somali QA API is running!"}