parse context

app.py

@@ -76,17 +76,22 @@ def add_docs(user_id: str, docs: list[str]) -> int:
     )
     return len(docs)
 # ----- Qwen-chat prompt helper ---------------------------------------------
-def build_qwen_prompt(context: str, user_question: str) -> str:
-    """Return a string that follows Qwen-Chat’s template."""
-    load_chat()                       # ← make sure tokenizer is ready
-    if context != "":
-        context = f"Context:\n{context}\n\n"
+def build_qwen_prompt(system: str, context: list[str], user_question: str) -> str:
+    """Return a Qwen-style prompt with multiple context items."""
+    load_chat()  # 確保 tokenizer 載入
+
     conversation = [
-        {"role": "system",
-         "content": "You are an email assistant. Use ONLY the context provided."},
-        {"role": "user",
-         "content": f"{context}{user_question}"}
+        {"role": "system", "content": system}
     ]
+
+    # 將每段 context 當作 user 提供的提示
+    for ctx in context:
+        if ctx.strip():  # 跳過空內容
+            conversation.append({"role": "user", "content": ctx})
+
+    # 加入最終問題
+    conversation.append({"role": "user", "content": user_question})
+
     return tokenizer.apply_chat_template(
         conversation, tokenize=False, add_generation_prompt=True
     )
@@ -103,7 +108,7 @@ def store_doc(doc_text: str, user_id="demo"):
     except Exception as e:
         return f"Error during storing: {e}"
 
-def answer(question: str, user_id="demo", history="None"):
+def answer(system: str, context: list[str], question: str, user_id="demo", history="None"):
     """UI callback: retrieve, build prompt with Qwen tags, generate answer."""
     try:
         if not question.strip():
@@ -111,7 +116,6 @@ def answer(question: str, user_id="demo", history="None"):
         if history != "None" and not kb[user_id]["texts"]:
             return "No reference passage yet. Add one first."
 
-        context = ""
         # 1.  Retrieve top-k similar passages
         if history == "Some":
             q_vec  = embed(question)
@@ -119,13 +123,13 @@ def answer(question: str, user_id="demo", history="None"):
             sims   = torch.matmul(store["vecs"], q_vec)          # [N]
             k      = min(4, sims.numel())
             idxs   = torch.topk(sims, k=k).indices.tolist()
-            context = "\n".join(store["texts"][i] for i in idxs)
+            context += [store["texts"][i] for i in idxs]
         elif history == "All":
             store  = kb[user_id]
-            context = "\n".join(store["texts"])
+            context += store["texts"]
 
         # 2.  Build a Qwen-chat prompt (helper defined earlier)
-        prompt = build_qwen_prompt(history + context, question)
+        prompt = build_qwen_prompt(system, context, question)
 
         # 3.  Generate and strip everything before the assistant tag
         load_chat()
@@ -168,16 +172,21 @@ with gr.Blocks() as demo:
 
     # ---- Q & A ----
     question_box = gr.Textbox(lines=2, label="Ask a question")
-    history_cb   = gr.Textbox(value="None", label="Use chat history")
+    history_cb = gr.Textbox(value="None", label="Use chat history")
+    system_box = gr.Textbox(lines=2, label="System prompt")
+    context_box = gr.Textbox(lines=6, label="Context passages (each line one passage)")
 
     answer_btn   = gr.Button("Answer")
     answer_box   = gr.Textbox(lines=6, label="Assistant reply")
 
     answer_btn.click(
-        fn=answer,
-        inputs=[question_box, user_id_box, history_cb],
-        outputs=answer_box
-    )
+    fn=lambda sys, ctx, q, uid, h: answer(sys,
+        [line.strip() for line in ctx.splitlines() if line.strip()],
+        q, uid, h
+    ),
+    inputs=[system_box, context_box, question_box, user_id_box, history_cb],
+    outputs=answer_box
+)
 
 # ---------- 3. FastAPI layer --------------------------------------------------
 class IngestReq(BaseModel):