better tracing

app.py

@@ -25,6 +25,7 @@ CTX_TOK_MAX    = 750          # sparar marginal till fråga + svar
 MAX_NEW_TOKENS = 128
 K              = 3
 DEFAULT_TEMP   = 0.3
+GEN_TIMEOUT    = 180  # Timeout för generering i sekunder
 
 # ── LOGGING ──────────────────────────────────────────────────────────
 logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
@@ -87,13 +88,20 @@ def chat_fn(q, temp, max_new_tokens, k, ctx_tok_max, history):
     history = history or []
     history.append({"role": "user", "content": q})
 
-    retriever = vs.as_retriever(search_kwargs={"k": int(k)})
-    docs = retriever.invoke(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
 
-    chunk_info = "\n\n".join([f"{i+1}. ({d.metadata['source']}) {d.page_content[:160]}…" for i, d in enumerate(docs)])
+    # 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):
@@ -132,9 +140,9 @@ def chat_fn(q, temp, max_new_tokens, k, ctx_tok_max, history):
     try:
         with concurrent.futures.ThreadPoolExecutor() as executor:
             future = executor.submit(generate)
-            ans = future.result(timeout=180)  # Timeout in seconds
+            ans = future.result(timeout=GEN_TIMEOUT)  # Timeout in seconds
     except concurrent.futures.TimeoutError:
-        ans = "⏰ Ingen respons från modellen inom 3 minuter."
+        ans = "⏰ 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}"
@@ -161,12 +169,14 @@ with gr.Blocks() as demo:
         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])
     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=[b_send]) 
 
 if __name__ == "__main__":
     demo.launch(share=True)   # ta bort share=True om du vill hålla den privat