adjust inputs

app.py

@@ -82,10 +82,12 @@ 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, history):
+def chat_fn(q, temp, max_new_tokens, k, ctx_tok_max, history):
     history = history or []
     history.append({"role": "user", "content": q})
 
+    # Skapa en ny retriever med valt k
+    retriever = vs.as_retriever(search_kwargs={"k": int(k)})
     docs = retriever.invoke(q)
     if not docs:
         history.append({"role": "assistant", "content": "🚫 Hittade inget relevant."})
@@ -95,16 +97,34 @@ def chat_fn(q, temp, history):
     chunk_info = "\n\n".join([f"{i+1}. ({d.metadata['source']}) {d.page_content[:160]}…" for i, d in enumerate(docs)])
     history.append({"role": "system", "content": f"🔎 Chunkar som används:\n{chunk_info}"})
 
-    prompt = build_prompt(q, docs)
-    history.append({"role": "system", "content": f"📝 Prompt till LLM:\n{prompt[:500]}{'...' if len(prompt)>500 else ''}"})  # Visa första 500 tecken
-
-    log.info(f"Prompt tokens={len(tokenizer.encode(prompt))}  temp={temp}")
+    # Bygg prompt med valt ctx_tok_max
+    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}")
 
     try:
         ans = gen_pipe(
             prompt,
             temperature=float(temp),
-            max_new_tokens=MAX_NEW_TOKENS,
+            max_new_tokens=int(max_new_tokens),
             pad_token_id=tokenizer.eos_token_id,
             eos_token_id=tokenizer.eos_token_id,
             do_sample=True,