app.py

import json, faiss, re
import numpy as np
import gradio as gr
from sentence_transformers import SentenceTransformer, CrossEncoder
from transformers import AutoTokenizer, AutoModelForQuestionAnswering, pipeline

INDEX_PATH = "e5_index.faiss"
TEXTS_PATH = "texts.json"

EMB_NAME = "intfloat/multilingual-e5-base"
READER_NAME = "savasy/bert-base-turkish-squad"
RERANK_MODEL = "BAAI/bge-reranker-v2-m3"

index = faiss.read_index(INDEX_PATH)
with open(TEXTS_PATH, "r", encoding="utf-8") as f:
    texts = json.load(f)

embedder = SentenceTransformer(EMB_NAME)
qa_tok = AutoTokenizer.from_pretrained(READER_NAME)
qa_mod = AutoModelForQuestionAnswering.from_pretrained(READER_NAME)
qa = pipeline("question-answering", model=qa_mod, tokenizer=qa_tok, device_map="auto")
reranker = CrossEncoder(RERANK_MODEL)

def search_semantic(q, k=80):
    qv = embedder.encode([f"query: {q}"], convert_to_numpy=True, normalize_embeddings=True)
    D, I = index.search(qv.astype(np.float32), k)
    return I[0].tolist()

def finalize_answer(raw_answer, context, max_chars=220):
    ans = (raw_answer or "").strip()
    if not ans:
        return "Bilmiyorum"
    sents = re.split(r"(?<=[.!?])\s+", context)
    hit = next((s for s in sents if ans in s), None)
    text = (hit or ans).strip()
    text = re.sub(r"\s+", " ", text).strip()
    if len(text) > max_chars:
        text = text[:max_chars].rsplit(" ", 1)[0].rstrip() + "…"
    if not re.search(r"[.!?…]$", text):
        text += "."
    return text[0].upper() + text[1:]

# ---- Cevaplayıcılar (kaynak döndürmez) ----
def answer_rerank(q, k=80, top_m=12, min_conf=0.10):
    I = search_semantic(q, k=k)
    cand = [texts[i] for i in I]
    scores = reranker.predict([[q, c] for c in cand])
    reranked = [x for _, x in sorted(zip(scores, cand), key=lambda z: z[0], reverse=True)]

    best = {"answer": "Bilmiyorum", "score": 0.0, "context": ""}
    for ctx in reranked[:top_m]:
        out = qa({"question": q, "context": ctx})
        ans, sc = out.get("answer", "").strip(), float(out.get("score", 0.0))
        if sc > best["score"]:
            best = {"answer": (ans or "Bilmiyorum"), "score": sc, "context": ctx}

    if best["score"] < float(min_conf) or not best["answer"]:
        best["answer"] = "Bilmiyorum"

    pretty = finalize_answer(best["answer"], best["context"])
    return pretty, round(best["score"], 3)

def answer_simple(q, k=80, top_m=12, min_conf=0.10):
    I = search_semantic(q, k=k)
    best = {"answer": "Bilmiyorum", "score": 0.0, "context": ""}
    for j in I[:top_m]:
        ctx = texts[j]
        out = qa({"question": q, "context": ctx})
        ans, sc = out.get("answer", "").strip(), float(out.get("score", 0.0))
        if sc > best["score"]:
            best = {"answer": (ans or "Bilmiyorum"), "score": sc, "context": ctx}
    if best["score"] < float(min_conf) or not best["answer"]:
        best["answer"] = "Bilmiyorum"
    pretty = finalize_answer(best["answer"], best["context"])
    return pretty, round(best["score"], 3)

# ---- Gradio UI (kaynak çıktısı ve ipucu yok) ----
with gr.Blocks() as demo:
    gr.Markdown("# Türkçe RAG QA (e5 + BERT-SQuAD)")
    inp = gr.Textbox(label="Sorunuzu yazın", placeholder="Verem nedir?")
    with gr.Row():
        k = gr.Slider(20, 120, value=80, step=10, label="k (retrieval)")
        m = gr.Slider(4, 24, value=12, step=1, label="top_m_for_qa")
        th = gr.Slider(0.05, 0.35, value=0.10, step=0.01, label="min_conf")
    use_rerank = gr.Checkbox(value=True, label="Reranker kullan (BGE v2 m3)")
    out_ans = gr.Textbox(label="Cevap")
    out_sc  = gr.Number(label="Güven", precision=3)
    btn = gr.Button("Sor")

    def route(q, kk, mm, tt, rr):
        if rr:
            return answer_rerank(q, kk, mm, tt)
        else:
            return answer_simple(q, kk, mm, tt)

    btn.click(route, inputs=[inp, k, m, th, use_rerank], outputs=[out_ans, out_sc])

demo.queue().launch()