Upload app.py

app.py

@@ -203,7 +203,7 @@ def get_tax_law(tax_type):
     }
     return tax_law_dict.get(tax_type, "無稅法")
 
-def fetch_law_summary(tax_law, keywords):
+def fetch_law_summary(query, tax_law, keywords):
     url = "https://ttc.mof.gov.tw/Api/GetData"
     headers = {
         "Content-Type": "application/x-www-form-urlencoded; charset=UTF-8",
@@ -274,12 +274,65 @@ def fetch_law_summary(tax_law, keywords):
             if tax_sn not in unique_results:
                 unique_results[tax_sn] = result
         for index, result in enumerate(list(unique_results.values())[:20]):  # 限制為前20個唯一結果
-            summary += f"""
-            <details>
-                <summary style="cursor: pointer; color: #0066cc;">{result['Title']}</summary>
-                <p>{result['Content']}</p>
-            </details>
+            client = OpenAI(
+                api_key=os.environ.get("YOUR_API_TOKEN"),
+                base_url="https://api.sambanova.ai/v1",
+            )
+            system_prompt = f"""
+            請判斷以下函釋內容與user提問的內容。
+            請給出一個0到100之間的相關性百分比，不要任何說明或理由。
+            回答格式為：
+            相關性：XX%
+
+            函釋內容：```{result['Content']}```
             """
+            prompt = f"""```{query}```"""
+            max_retries = 2
+            retry_delay = 4
+            for attempt in range(max_retries):
+                try:
+                    response = client.chat.completions.create(
+                    model="Meta-Llama-3.1-405B-Instruct",
+                    messages=[
+                            {"role": "system", "content": system_prompt},
+                            {"role": "user", "content": prompt}
+                        ],
+                        temperature=0.7
+                    )
+                    relevance_percentage = response.choices[0].message.content.strip()
+                    break
+                except Exception as e:
+                    if (attempt == 0) or (attempt == max_retries - 1):
+                        try:
+                            response = client.chat.completions.create(
+                                model="Meta-Llama-3.1-70B-Instruct",
+                                messages=[
+                                    {"role": "system", "content": system_prompt},
+                                    {"role": "user", "content": prompt}
+                                ],
+                                temperature=0.7
+                            )
+                            relevance_percentage = response.choices[0].message.content.strip()
+                            break
+                        except Exception as e2:
+                            relevance_percentage = "相關性：0%"
+                            break
+                    else:
+                        print(f"Retrying in {retry_delay} seconds...")
+                        time.sleep(retry_delay)
+                        retry_delay *= 2
+            try:
+                percentage = int(relevance_percentage.split("：")[1].strip().rstrip('%'))
+            except ValueError:
+                print(f"Warning: Could not parse relevance percentage from '{relevance_percentage}'")
+                percentage = 0
+            if percentage > 0:
+                summary += f"""
+                <details>
+                    <summary style="cursor: pointer; color: #0066cc;">{result['Title']} （相關性：{percentage} %）</summary>
+                    <p>{result['Content']}</p>
+                </details>
+                """
         return summary
     else:
         return "<p>未檢索到相關法令彙編函釋。</p>"
@@ -305,7 +358,7 @@ def llm_openai_api(query, answer):
                 {"role": "user", "content": user_prompt}
             ],
             temperature=0.7,
-            top_p=1
+            top_p=0.9
         )
         return response.choices[0].message.content.strip()
     except Exception as e:
@@ -323,8 +376,8 @@ def handle_interaction(query, api_key, state):
         params: resend.Emails.SendParams = {
             "from": "Tax_KM <[email protected]>",
             "to": ["[email protected]"],
-            "subject": "地方稅知識庫系統 API KEY",
-            "html": f"<strong>查詢內容：<br>{query}</strong>",
+            "subject": "地方稅知識庫 API KEY",
+            "html": f"<strong>檢索內容：<br>{query}</strong>",
         }
         try:
             email_response = resend.Emails.send(params)
@@ -349,7 +402,7 @@ def handle_interaction(query, api_key, state):
             print(f"Tax Law: {tax_law}")
             print(f"Keywords: {keywords}")
         tax_law = get_tax_law(tax_name)
-        law_summary_content = fetch_law_summary(tax_law, keywords)
+        law_summary_content = fetch_law_summary(query, tax_law, keywords)
     else:
         law_summary_content = ""
         gr.Info(f"多個問題不會提供法令彙編函釋檢索結果。")
@@ -455,8 +508,8 @@ with gr.Blocks(theme=gr.themes.Soft(), css=custom_css) as iface:
     > ### **※ RAG-based KM 以地方稅極少知識資料作示範，僅供參考，準確資訊請依地方稅稽徵機關回覆為準。系統部署：江信宗，LLM：Llama-3.1-70B。**
     """, elem_classes="center-text")
     with gr.Row():
-        query_input = gr.Textbox(label="輸入您的問題，系統將基於學習到的知識資料提供相關答案。", placeholder="請輸入您的問題（支援同時輸入多個問題，例如：問題1？問題2？）", autofocus=True, scale=3, max_lines=5, elem_classes="query-input")
-        api_key_input = gr.Textbox(label="請輸入您的 API Key", type="password", placeholder="API authentication key", scale=1, elem_classes="api-key-input")
+        query_input = gr.Textbox(label="輸入您的問題，系統將基於學習到的知識資料提供相關答案。", placeholder="支援同時輸入多個問題，例如：問題1？問題2？", autofocus=True, scale=3, max_lines=5, elem_classes="query-input")
+        api_key_input = gr.Textbox(label="輸入您的 API Key", type="password", placeholder="API authentication key", scale=1, elem_classes="api-key-input")
     answer_output = gr.Textbox(label="知識庫答案", interactive=False, max_lines=40, elem_classes="answer-box")
     with gr.Row():
         insight_q1 = gr.Button("洞察問題 1", visible=False, elem_classes=["insight-btn"])