update 42

app.py

@@ -1,181 +1,148 @@
-import os
-import logging
-
+# ── app.py ─────────────────────────────────────────────────────────
+import os, logging, math, textwrap
 import gradio as gr
 from transformers import pipeline
 from langchain_community.document_loaders import PyPDFLoader
 from langchain_community.vectorstores import FAISS
 from langchain_huggingface.embeddings import HuggingFaceEmbeddings
-from langchain_huggingface.llms import HuggingFacePipeline
-from langchain.chains import RetrievalQA
 from langchain.text_splitter import RecursiveCharacterTextSplitter
 
-# ── KONFIGURATION ───────────────────────────────────────────
+# ── KONFIGURATION ─────────────────────────────────────────
 DOCS_DIR       = "document"
 INDEX_DIR      = "faiss_index"
-CHUNK_SIZE     = 500
-CHUNK_OVERLAP  = 50
 EMB_MODEL      = "KBLab/sentence-bert-swedish-cased"
 LLM_MODEL      = "tiiuae/falcon-rw-1b"
-MAX_NEW_TOKENS = 64
+
+CHUNK_SIZE     = 500
+CHUNK_OVERLAP  = 50
+MAX_NEW_TOKENS = 128          # svarlängd
+CTX_TOKEN_MAX  = 900          # för att stanna under modell‑begränsningen 1 024
+K              = 10           # hur många chunkar vi hämtar
 DEFAULT_TEMP   = 0.3
-K              = 10
 
-# ── LOGGING ─────────────────────────────────────────────────
-logging.basicConfig(
-    level=logging.INFO,
-    format="%(asctime)s %(levelname)s %(message)s"
-)
+# ── LOGGING ──────────────────────────────────────────────
+logging.basicConfig(level=logging.INFO, format="%(asctime)s  %(levelname)s  %(message)s")
 logger = logging.getLogger(__name__)
 
-# ── 1) Bygg eller ladda FAISS-index ─────────────────────────
+# ── 1) Index (bygg eller ladda) ─────────────────────────
 emb = HuggingFaceEmbeddings(model_name=EMB_MODEL)
+
 if os.path.isdir(INDEX_DIR):
-    logger.info(f"🔄 Laddar befintligt FAISS-index från `{INDEX_DIR}`…")
+    logger.info(f"🔄 Laddar FAISS‑index från `{INDEX_DIR}`")
     vs = FAISS.load_local(INDEX_DIR, emb)
 else:
-    logger.info("⚙️ Bygger nytt FAISS-index från PDF-filer…")
-    splitter = RecursiveCharacterTextSplitter(
-        chunk_size=CHUNK_SIZE, chunk_overlap=CHUNK_OVERLAP
-    )
-    all_docs, files = [], []
+    logger.info("⚙️ Bygger FAISS‑index från PDF‑filer …")
+    splitter = RecursiveCharacterTextSplitter(chunk_size=CHUNK_SIZE,
+                                              chunk_overlap=CHUNK_OVERLAP)
+    docs, files = [], []
     for fn in os.listdir(DOCS_DIR):
         if fn.lower().endswith(".pdf"):
-            path   = os.path.join(DOCS_DIR, fn)
-            loader = PyPDFLoader(path)
-            pages  = loader.load()
+            pages  = PyPDFLoader(os.path.join(DOCS_DIR, fn)).load()
             chunks = splitter.split_documents(pages)
             for c in chunks:
                 c.metadata["source"] = fn
-            all_docs.extend(chunks)
+            docs.extend(chunks)
             files.append(fn)
-    vs = FAISS.from_documents(all_docs, emb)
+    vs = FAISS.from_documents(docs, emb)
     vs.save_local(INDEX_DIR)
-    logger.info(f"✅ Sparade index i `{INDEX_DIR}` ({len(files)} PDF, {len(all_docs)} chunkar)")
-
-# ── 2) Initiera LLM ──────────────────────────────────────────
-logger.info("🚀 Initierar LLM-pipeline…")
-pipe = pipeline(
-    "text-generation",
-    model=LLM_MODEL,
-    device=-1,
-    max_new_tokens=MAX_NEW_TOKENS
-)
-llm = HuggingFacePipeline(
-    pipeline=pipe,
-    model_kwargs={
-        "temperature": DEFAULT_TEMP,
-        "max_new_tokens": MAX_NEW_TOKENS
-    }
-)
-logger.info("✅ LLM-pipeline initierad")
-
-# ── 3) Bygg RetrievalQA ─────────────────────────────────────
+    logger.info(f"✅ Sparade index ({len(files)} PDF, {len(docs)} chunkar)")
+
 retriever = vs.as_retriever(search_kwargs={"k": K})
-qa_chain  = RetrievalQA.from_chain_type(
-    llm=llm,
-    retriever=retriever,
-    return_source_documents=False,
-    chain_type="stuff"
-)
-logger.info(f"✅ RetrievalQA klar (k={K})")
-
-# ── 4) Gradio-funktioner ────────────────────────────────────
-def ping() -> str:
-    logger.debug("ping() anropad")
-    return "pong"
+
+# ── 2) LLM‑pipeline ─────────────────────────────────────
+logger.info("🚀 Initierar text‑genererings‑pipeline …")
+gen_pipe = pipeline("text-generation",
+                    model=LLM_MODEL,
+                    device=-1,
+                    max_new_tokens=MAX_NEW_TOKENS)
+
+logger.info("✅ LLM klar")
+
+# ── 3) Hjälpfunktioner ──────────────────────────────────
+def truncate_tokens(text: str, max_tokens: int = CTX_TOKEN_MAX) -> str:
+    """Väldigt enkel token‑approx (1 token ≈4 tecken)"""
+    approx_tokens = len(text) // 4
+    if approx_tokens <= max_tokens:
+        return text
+    slice_len = max_tokens * 4
+    return text[:slice_len]
 
 def test_retrieval(query: str) -> str:
-    logger.debug(f"test_retrieval() anropad med query={query!r}")
     docs = retriever.get_relevant_documents(query)
     if not docs:
         return "🚫 Inga träffar"
     out = []
-    for i, d in enumerate(docs, start=1):
-        src = d.metadata.get("source", "okänd")
-        snippet = d.page_content.replace("\n", " ")[:200]
+    for i, d in enumerate(docs, 1):
+        src     = d.metadata.get("source", "okänd")
+        snippet = d.page_content.replace("\n", " ")[:160]
         out.append(f"{i}. ({src}) …{snippet}…")
     return "\n\n".join(out)
 
-def chat_fn(
-    query: str,
-    temperature: float,
-    history: list[dict]
-) -> tuple[list[dict], list[dict]]:
-    """
-    query: användarens fråga
-    temperature: slump-parameter för LLM
-    history: tidigare chatt-meddelanden som lista av {"role","content"}
-    return: (uppdaterad_history, uppdaterad_history)
-    """
-    logger.info(f"chat_fn() anropad med query={query!r}, temp={temperature}, history_len={len(history) if history else 0}")
+# ── 4) Chat‑funktionen (exakt 3 param, 2 retur) ─────────
+def chat_fn(query: str, temperature: float, history: list[dict]):
     history = history or []
+    history.append({"role": "user", "content": query})
 
-    if not isinstance(query, str):
-        msg = f"❌ Fel: query måste vara str, fick {type(query)}"
-        logger.error(msg)
-        history.append({"role": "assistant", "content": msg})
+    # Hämtar K chunkar
+    docs = retriever.get_relevant_documents(query)
+    if not docs:
+        ans = "🚫 Hittade inget relevant innehåll i dokumenten."
+        history.append({"role": "assistant", "content": ans})
         return history, history
 
-    if not query.strip():
-        msg = "⚠️ Du måste skriva en fråga."
-        logger.warning(msg)
-        history.append({"role": "assistant", "content": msg})
-        return history, history
+    # Bygg kontext och trimma
+    context = "\n\n---\n\n".join(d.page_content for d in docs)
+    context = truncate_tokens(context, CTX_TOKEN_MAX)
 
-    history.append({"role": "user", "content": query})
+    prompt = textwrap.dedent(f"""
+        Du är en hjälpsam assistent som svarar på svenska.
+        Kontext (ur PDF‑dokument):
 
-    # Bygg prompt
-    docs    = retriever.get_relevant_documents(query)
-    context = "\n\n---\n\n".join(d.page_content for d in docs)
-    prompt  = f"Kontekst:\n{context}\n\nFråga: {query}"
-    logger.debug(f"Prompt till QA-kedjan: {prompt[:200]}...")
+        {context}
+
+        Fråga: {query}
+        Svar (svenska):""").strip()
+
+    logger.info(f"📝 Prompt‑längd≈{len(prompt)} tecken, temp={temperature}")
 
-    # Justera temp
-    try:
-        llm.pipeline.model_kwargs["temperature"] = temperature
-    except Exception as e:
-        logger.exception("Kunde inte sätta temperature")
-    
     try:
-        svar = qa_chain.run(prompt)
+        resp = gen_pipe(prompt,
+                        temperature=float(temperature),
+                        max_new_tokens=MAX_NEW_TOKENS,
+                        pad_token_id=2,
+                        eos_token_id=2,
+                        do_sample=True,
+                        return_full_text=False)[0]["generated_text"]
     except Exception as e:
-        logger.exception("Fel vid anrop av qa_chain")
-        svar = f"❌ Fel vid QA-kedjan: {e}"
-
-    # Ange källa
-    src = docs[0].metadata.get("source", "okänd") if docs else "okänd"
-    content = f"**(Dokument: {src})**\n\n{svar}"
-    history.append({"role": "assistant", "content": content})
-    logger.info(f"chat_fn färdig, svarslängd={len(svar)} tecken, totalt history={len(history)}")
+        logger.exception("Fel vid generering")
+        resp = f"❌ Fel: {e}"
 
+    src_hint = docs[0].metadata.get("source", "okänd")
+    history.append({"role": "assistant",
+                    "content": f"**(Källa: {src_hint})**\n\n{resp}"})
     return history, history
 
-# ── 5) Bygg Gradio-UI ────────────────────────────────────────
+# ── 5) Bygg Gradio‑UI ──────────────────────────────────
 with gr.Blocks() as demo:
-    gr.Markdown("# 🌟 Svensk Dokumentassistent")
-    gr.Markdown(f"**✅ Index laddat från** `{INDEX_DIR}`")
+    gr.Markdown("# 📚 svensk RAG‑chat\nStäll frågor till dina PDF‑filer")
+    gr.Markdown(f"**PDF‑filer i index:** {', '.join(os.listdir(DOCS_DIR)) or 'inga'}")
 
     with gr.Row():
-        ping_btn = gr.Button("🏓 Ping")
-        ping_out = gr.Textbox(label="Ping-svar")
-
-        test_btn = gr.Button("🔎 Testa Retrieval")
-        test_out = gr.Textbox(label="Retrieval-snippet")
+        test_in  = gr.Textbox(label="Snabb‑retrieval (ingen AI)", lines=1)
+        test_btn = gr.Button("🔎 Testa")
+        test_out = gr.Textbox(label="Chunkar")
 
     with gr.Row():
-        txt   = gr.Textbox(lines=2, placeholder="Ex: Vad är krav?", label="Din fråga")
-        temp  = gr.Slider(0.0, 1.0, value=DEFAULT_TEMP, step=0.05, label="Temperatur")
-        send  = gr.Button("📨 Skicka")
+        q_in   = gr.Textbox(placeholder="Ex: Vad står det om krav?", label="Fråga")
+        temp   = gr.Slider(0, 1, value=DEFAULT_TEMP, step=0.05, label="Temperatur")
+        send   = gr.Button("📨 Skicka")
 
-    chatbot    = gr.Chatbot(type="messages", label="Chatt")
-    chat_state = gr.State([])
+    chat      = gr.Chatbot(type="messages", label="Chat")
+    chat_hist = gr.State([])
 
-    ping_btn.click(fn=ping, inputs=[], outputs=[ping_out])
-    test_btn.click(fn=test_retrieval, inputs=[txt], outputs=[test_out])
-    send.click(fn=chat_fn, inputs=[txt, temp, chat_state], outputs=[chatbot, chat_state])
+    # Kopplingar
+    test_btn.click(test_retrieval, inputs=[test_in], outputs=[test_out])
+    send.click(chat_fn, inputs=[q_in, temp, chat_hist], outputs=[chat, chat_hist])
 
 if __name__ == "__main__":
-    # share=True om du vill publikt dela länken
     demo.launch(share=True)
-