app.py

# ── app.py ───────────────────────────────────────────────────────────
import os, logging, textwrap
import gradio as gr
from transformers import pipeline, AutoTokenizer
from langchain_community.document_loaders import PyPDFLoader
from langchain_community.vectorstores import FAISS
from langchain_huggingface.embeddings import HuggingFaceEmbeddings
from langchain.text_splitter import RecursiveCharacterTextSplitter
import concurrent.futures

# ── KONFIG ───────────────────────────────────────────────────────────
DOCS_DIR       = "document"
INDEX_DIR      = "faiss_index"
EMB_MODEL      = "KBLab/sentence-bert-swedish-cased"
#LLM_MODEL      = "tiiuae/falcon-rw-1b" # Dålig
#LLM_MODEL      = "google/flan-t5-base" # Dålig
#LLM_MODEL      = "bigscience/bloom-560m" # Dålig
#LLM_MODEL      = "NbAiLab/nb-gpt-j-6B" #- Restricted 
#LLM_MODEL      = "datificate/gpt2-small-swedish" # Finns ej på Hugging face
#LLM_MODEL      =  "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
# timpal0l/mdeberta-v3-base-squad2 liten och möjlig på Svenska
#LLM_MODEL      = "AI-Sweden-Models/gpt-sw3-1.3B" # finns olika varianter 126M, 356M, 1.3B, 6.7B, 20B, 40B
LLM_MODEL      = "AI-Sweden-Models/gpt-sw3-356M"
# LLM_MODEL = AI-Sweden-Models/Llama-3-8B-instruct # kanske för stor
# https://www.ai.se/en/ai-labs/natural-language-understanding/models-resources

CHUNK_SIZE     = 400
CHUNK_OVERLAP  = 40
CTX_TOK_MAX    = 750          # sparar marginal till fråga + svar
MAX_NEW_TOKENS = 512
K              = 5
DEFAULT_TEMP   = 0.8
GEN_TIMEOUT    = 180  # Timeout för generering i sekunder

# ── LOGGING ──────────────────────────────────────────────────────────
logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
log = logging.getLogger(__name__)

# ── 1) Index (bygg eller ladda) ─────────────────────────────────────
emb = HuggingFaceEmbeddings(model_name=EMB_MODEL)
INDEX_PATH = os.path.join(INDEX_DIR, "index.faiss")
if os.path.isfile(INDEX_PATH):
    log.info(f"🔄 Laddar index från {INDEX_DIR}")
    vs = FAISS.load_local(INDEX_DIR, emb)
else:
    splitter = RecursiveCharacterTextSplitter(chunk_size=CHUNK_SIZE, chunk_overlap=CHUNK_OVERLAP)
    docs, pdfs = [], []
    for fn in os.listdir(DOCS_DIR):
        if fn.lower().endswith(".pdf"):
            chunks = splitter.split_documents(PyPDFLoader(os.path.join(DOCS_DIR, fn)).load())
            for c in chunks:
                c.metadata["source"] = fn
            docs.extend(chunks); pdfs.append(fn)
    vs = FAISS.from_documents(docs, emb); vs.save_local(INDEX_DIR)
    log.info(f"✅ Byggt index – {len(pdfs)} PDF / {len(docs)} chunkar")
retriever = vs.as_retriever(search_kwargs={"k": K})

# ── 2) LLM‑pipeline & tokenizer ─────────────────────────────────────
log.info("🚀 Initierar LLM …")
gen_pipe  = pipeline("text-generation", model=LLM_MODEL, device=-1, max_new_tokens=MAX_NEW_TOKENS)
tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL)
log.info("✅ LLM klar")

# ── 3) Hjälpfunktioner ──────────────────────────────────────────────
def build_prompt(query: str, docs):
    """
    Tar så många chunkar som ryms i CTX_TOK_MAX token
    """
    context_parts = []
    total_ctx_tok = 0
    for d in docs:
        tok_len = len(tokenizer.encode(d.page_content))
        if total_ctx_tok + tok_len > CTX_TOK_MAX:
            break
        context_parts.append(d.page_content)
        total_ctx_tok += tok_len

    context = "\n\n---\n\n".join(context_parts)
    return textwrap.dedent(f"""\
        Du är en hjälpsam assistent som svarar på svenska.
        Kontext (hämtat ur PDF‑dokument):

        {context}

        Fråga: {query}
        Svar (svenska):""").strip()

def test_retrieval(q):  # snabb‑test utan AI
    docs = retriever.invoke(q)
    return "\n\n".join([f"{i+1}. ({d.metadata['source']}) {d.page_content[:160]}…" for i, d in enumerate(docs)]) or "🚫 Inga träffar"

def chat_fn(q, temp, max_new_tokens, k, ctx_tok_max, history):
    history = history or []
    history.append({"role": "user", "content": q})

    # Hämta chunkar och poäng
    docs_and_scores = vs.similarity_search_with_score(q, k=int(k))
    docs = [doc for doc, score in docs_and_scores]
    scores = [score for doc, score in docs_and_scores]

    if not docs:
        history.append({"role": "assistant", "content": "🚫 Hittade inget relevant."})
        return history, history

    # Visa chunkar och poäng
    chunk_info = "\n\n".join([
        f"{i+1}. ({d.metadata['source']}) score={scores[i]:.3f}\n{d.page_content[:160]}…"
        for i, d in enumerate(docs)
    ])
    history.append({"role": "system", "content": f"🔎 Chunkar som används:\n{chunk_info}"})

    def build_prompt_dynamic(query, docs, ctx_tok_max):
        context_parts = []
        total_ctx_tok = 0
        for d in docs:
            tok_len = len(tokenizer.encode(d.page_content))
            if total_ctx_tok + tok_len > int(ctx_tok_max):
                break
            context_parts.append(d.page_content)
            total_ctx_tok += tok_len
        context = "\n\n---\n\n".join(context_parts)
        return textwrap.dedent(f"""\
            Du är en hjälpsam assistent som svarar på svenska.
            Kontext (hämtat ur PDF‑dokument):

            {context}

            Fråga: {query}
            Svar (svenska):""").strip()

    prompt = build_prompt_dynamic(q, docs, ctx_tok_max)
    log.info(f"Prompt tokens={len(tokenizer.encode(prompt))}  temp={temp}  max_new_tokens={max_new_tokens} k={k} ctx_tok_max={ctx_tok_max}")

    def generate():
        return gen_pipe(
            prompt,
            temperature=float(temp),
            max_new_tokens=int(max_new_tokens),
            pad_token_id=tokenizer.eos_token_id,
            eos_token_id=tokenizer.eos_token_id,
            do_sample=True,
            return_full_text=False
        )[0]["generated_text"]

    try:
        with concurrent.futures.ThreadPoolExecutor() as executor:
            future = executor.submit(generate)
            ans = future.result(timeout=GEN_TIMEOUT)  # Timeout in seconds
    except concurrent.futures.TimeoutError:
        ans = f"⏰ Ingen respons från modellen inom {GEN_TIMEOUT} sekunder."
    except Exception as e:
        log.exception("Genererings‑fel")
        ans = f"❌ Fel vid generering: {type(e).__name__}: {e}\n\nPrompt:\n{prompt}"

    src_hint = docs[0].metadata["source"] if docs else "Ingen källa"
    history.append({"role": "assistant", "content": f"**(Källa: {src_hint})**\n\n{ans}"})
    return history, history

# ── 4) Gradio UI ────────────────────────────────────────────────────
with gr.Blocks() as demo:
    gr.Markdown("# 📚 Svensk RAG‑chat")
    gr.Markdown(f"**PDF‑filer:** {', '.join(os.listdir(DOCS_DIR)) or '–'}")
    gr.Markdown(f"**LLM-modell som används:** `{LLM_MODEL}`", elem_id="llm-info")

    with gr.Row():
        q_test = gr.Textbox(label="🔎 Test Retrieval")
        b_test = gr.Button("Testa")
        o_test = gr.Textbox(label="Chunkar")

    with gr.Row():
        q_in   = gr.Textbox(label="Fråga", placeholder="Ex: Vad är förvaltningsöverlämnande?")
        temp   = gr.Slider(0, 1, value=DEFAULT_TEMP, step=0.05, label="Temperatur")
        max_new_tokens = gr.Slider(32, 1024, value=MAX_NEW_TOKENS, step=8, label="Max svarslängd (tokens)")
        k      = gr.Slider(1, 10, value=K, step=1, label="Antal chunkar (K)")
        ctx_tok_max = gr.Slider(100, 2000, value=CTX_TOK_MAX, step=50, label="Max kontexttokens")
        b_send = gr.Button("Skicka")
        b_stop = gr.Button("Stoppa")  # Lägg till stoppknapp

    chat      = gr.Chatbot(type="messages", label="Chat")
    chat_hist = gr.State([])

    b_test.click(test_retrieval, inputs=[q_test], outputs=[o_test])
    send_event = b_send.click(
        chat_fn,
        inputs=[q_in, temp, max_new_tokens, k, ctx_tok_max, chat_hist],
        outputs=[chat, chat_hist]
    )
    b_stop.click(None, cancels=[send_event])

if __name__ == "__main__":
    demo.launch(share=True)   # ta bort share=True om du vill hålla den privat