App creada

app.py

@@ -0,0 +1,75 @@
+from fastapi import FastAPI, HTTPException
+from fastapi.middleware.cors import CORSMiddleware
+from pydantic import BaseModel
+from deep_translator import GoogleTranslator
+from datasets import load_dataset
+from transformers import TapexTokenizer, BartForConditionalGeneration
+import pandas as pd, torch, os
+
+# === Config ===
+HF_MODEL_ID = os.getenv("HF_MODEL_ID", "stvnnnnnn/tapex-wikisql-best")
+TABLE_SPLIT = os.getenv("TABLE_SPLIT", "validation")
+TABLE_INDEX = int(os.getenv("TABLE_INDEX", "10"))
+MAX_ROWS    = int(os.getenv("MAX_ROWS", "12"))
+
+torch.set_num_threads(1)
+
+app = FastAPI(title="NL→SQL – TAPEX + WikiSQL (HF Space)")
+
+# CORS: permite que Vercel (o cualquier origen) consuma la API
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"], allow_credentials=False,
+    allow_methods=["*"], allow_headers=["*"],
+)
+
+# Carga modelo/tokenizer con bajo pico de RAM (CPU)
+tok = TapexTokenizer.from_pretrained(HF_MODEL_ID)
+model = BartForConditionalGeneration.from_pretrained(
+    HF_MODEL_ID, low_cpu_mem_usage=True
+).to("cpu")
+
+class NLQuery(BaseModel):
+    nl_query: str
+
+def get_example(split, index):
+    # streaming para no cargar todo WikiSQL en RAM
+    ds = load_dataset("Salesforce/wikisql", split=split, streaming=True)
+    for i, ex in enumerate(ds):
+        if i == index:
+            return ex
+    raise IndexError("Index fuera de rango")
+
+def load_table(split=TABLE_SPLIT, index=TABLE_INDEX, max_rows=MAX_ROWS):
+    ex = get_example(split, index)
+    header = [str(h) for h in ex["table"]["header"]]
+    rows = ex["table"]["rows"][:max_rows]
+    return pd.DataFrame(rows, columns=header)
+
+@app.get("/api/health")
+def health():
+    return {"ok": True, "model": HF_MODEL_ID, "split": TABLE_SPLIT, "index": TABLE_INDEX}
+
+@app.get("/api/preview")
+def preview():
+    df = load_table()
+    return {"columns": df.columns.tolist(), "rows": df.head(8).to_dict(orient="records")}
+
+@app.post("/api/nl2sql")
+def nl2sql(q: NLQuery):
+    try:
+        text = (q.nl_query or "").strip()
+        if not text:
+            raise ValueError("Consulta vacía.")
+
+        is_ascii = all(ord(c) < 128 for c in text)
+        query_en = text if is_ascii else GoogleTranslator(source="auto", target="en").translate(text)
+
+        df = load_table()
+        enc = tok(table=df, query=query_en, return_tensors="pt", truncation=True)
+        out = model.generate(**enc, max_length=160, num_beams=1)
+        sql = tok.batch_decode(out, skip_special_tokens=True)[0]
+
+        return {"consulta_original": text, "consulta_traducida": query_en, "sql_generado": sql}
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))