rvdbra commited on
Commit
f176f72
·
verified ·
1 Parent(s): c106b8f

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +254 -133
app.py CHANGED
@@ -1,39 +1,80 @@
1
  import os
 
 
 
2
  from dotenv import load_dotenv
3
- load_dotenv() # loads variables from a local .env for local testing
4
 
5
- # 1) LLM and embeddings
6
- from langchain.chat_models import init_chat_model
7
- from langchain_openai import OpenAIEmbeddings
8
  from pinecone import Pinecone
9
- from langchain_pinecone import PineconeVectorStore
10
-
11
- # initialize OpenAI chat model
12
- llm = init_chat_model(
13
- os.getenv("OPENAI_API_KEY"),
14
- model_provider="openai",
15
- model="gpt-4.1-2025-04-14",
16
- temperature=0.0,
17
- )
18
 
19
- # embeddings + Pinecone vector store
20
- embeddings = OpenAIEmbeddings(model="text-embedding-3-large")
21
- pinecone_api_key = os.getenv("PINECONE_API_KEY")
22
- pinecone_index_name = os.getenv("PINECONE_INDEX_NAME")
23
- pc = Pinecone(api_key=pinecone_api_key)
24
- index = pc.Index(pinecone_index_name)
25
- vector_store = PineconeVectorStore(embedding=embeddings, index=index)
26
 
27
- # 2) Supabase logging
28
- from supabase import create_client, Client
29
- from datetime import datetime
 
 
 
 
30
 
31
- supabase: Client = create_client(
32
- os.getenv("SUPABASE_URL"),
33
- os.getenv("SUPABASE_KEY")
34
- )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
35
 
36
  def log_interaction_cloud(user_id: str, question: str, answer: str):
 
 
 
 
37
  payload = {
38
  "user_id": user_id,
39
  "question": question,
@@ -42,135 +83,215 @@ def log_interaction_cloud(user_id: str, question: str, answer: str):
42
  }
43
  try:
44
  res = supabase.table("logs").insert(payload).execute()
45
- if hasattr(res, "status_code") and res.status_code >= 400:
46
- print(f"[Supabase Error #{res.status_code}]:", res.data)
47
- else:
48
- print("[Supabase Success]:", res.data)
49
  except Exception as e:
50
- print("[Supabase Exception]", e)
51
 
52
- # 3) Conversation memory & user IDs
53
- import uuid
54
- from langchain.memory import ConversationBufferWindowMemory
55
  def generate_user_id():
56
  return "user_" + uuid.uuid4().hex[:6]
57
 
 
58
  memory = ConversationBufferWindowMemory(
59
  memory_key="chat_history",
60
  return_messages=True,
61
  k=5
62
  )
63
 
64
- # 4) Retrieval tool
65
- from langchain_core.tools import tool
66
-
67
  @tool(response_format="content_and_artifact")
68
  def retrieve(query: str):
69
- docs = vector_store.similarity_search(query, k=5)
70
- serialized = "\n\n".join(
71
- f"Source: {d.metadata}\nContent: {d.page_content}"
72
- for d in docs
73
- )
74
- return serialized, docs
75
-
76
- # 5) Build LangGraph
77
- from langgraph.graph import MessagesState, StateGraph, END
78
- from langgraph.prebuilt import ToolNode, tools_condition
79
- from langchain_core.messages import SystemMessage
 
 
80
 
81
  SYSTEM_MESSAGE_TEXT = (
82
- "You are “Einbürgerungs-Helfer” … [your long system prompt here] …"
 
 
 
83
  )
84
 
85
- def query_or_respond(state: MessagesState):
86
- sys_msg = SystemMessage(SYSTEM_MESSAGE_TEXT)
87
- raw_hist = memory.load_memory_variables({})["chat_history"]
88
- # filter off-topic if needed…
89
- prompt = [sys_msg] + raw_hist[-5:]
90
- llm_with_tools = llm.bind_tools([retrieve])
91
- resp = llm_with_tools.invoke(prompt)
92
- # log if direct
93
- user_q = [m.content for m in state["messages"] if m.type=="human"][-1]
94
- memory.save_context({"input": user_q}, {"output": resp.content})
95
- if not getattr(resp, "tool_calls", None):
96
- log_interaction_cloud(state.get("user_id",""), user_q, resp.content)
97
- return {"messages": state["messages"] + [resp]}
98
-
99
- tools_node = ToolNode([retrieve])
100
-
101
- def generate(state: MessagesState):
102
- # collect any tool outputs
103
- tool_msgs = []
104
- for m in reversed(state["messages"]):
105
- if m.type=="tool": tool_msgs.append(m)
106
- else: break
107
- docs = "\n\n".join(m.content for m in tool_msgs[::-1])
108
- prompt = [SystemMessage(SYSTEM_MESSAGE_TEXT + ("\n\n"+docs if docs else ""))]
109
- # build recent convo
110
- convo = []
111
- for m in state["messages"]:
112
- if m.type=="human" or (m.type=="ai" and not getattr(m,"tool_calls",None)):
113
- convo.append(m)
114
- prompt += convo[-6:]
115
- resp = llm.invoke(prompt)
116
- # log final
117
- user_q = [m.content for m in state["messages"] if m.type=="human"][-1]
118
- log_interaction_cloud(state.get("user_id",""), user_q, resp.content)
119
- return {"messages": state["messages"] + [resp]}
120
-
121
- g = StateGraph(MessagesState)
122
- g.add_node(query_or_respond)
123
- g.add_node(tools_node)
124
- g.add_node(generate)
125
- g.set_entry_point("query_or_respond")
126
- g.add_conditional_edges("query_or_respond", tools_condition, {END: END, "tools": "tools"})
127
- g.add_edge("tools", "generate")
128
- g.add_edge("generate", END)
129
- graph = g.compile()
130
-
131
- # 6) Gradio UI
132
- import gradio as gr
133
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
134
  chat_histories = {}
135
 
136
- def stream_response(user_input, user_id):
137
- if not user_id: user_id = "anonymous"
138
- state = {"messages":[{"role":"user","user_id":user_id,"content":user_input}], "user_id":user_id}
139
- collected = ""
140
- for step in graph.stream(state, stream_mode="values"):
141
- msg = step["messages"][-1].content
142
- for t in msg:
143
- collected += t
144
- yield collected
145
-
146
- def respond(user_msg, chat_history, user_id):
147
- if user_id not in chat_histories: chat_histories[user_id] = []
148
- gen = stream_response(user_msg, user_id)
149
- full = ""
150
- for partial in gen:
151
- full = partial
152
- yield chat_history + [[user_msg, full]]
153
- chat_histories[user_id].append([user_msg, full])
 
 
 
 
 
 
 
 
 
154
 
155
  def clear_history():
156
  return "", []
157
 
158
- description = """### Einbürgerungs-Helfer
159
- … your multilingual description …"""
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
160
 
161
- with gr.Blocks() as iface:
162
- gr.Markdown("# Einbürgerungs-Helfer\n" + description)
163
  with gr.Row():
164
- user_id_input = gr.Textbox(label="User ID", placeholder="z.B. user_123")
165
- clear_btn = gr.Button("🗑️ Clear")
166
- chatbot = gr.Chatbot(value=[[None,"Ich helfe…"]])
167
- msg_in = gr.Textbox(label="Your question")
168
- btn = gr.Button("Send")
169
- btn.click(respond, [msg_in, chatbot, user_id_input], chatbot)
170
- clear_btn.click(clear_history, [], [user_id_input, chatbot])
171
-
172
- iface.launch(
173
- server_name="0.0.0.0",
174
- server_port=int(os.getenv("PORT", 7860)),
175
- share=False
176
- )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  import os
2
+ import uuid
3
+ from datetime import datetime
4
+ from typing_extensions import TypedDict
5
  from dotenv import load_dotenv
 
6
 
7
+ import gradio as gr
8
+ from langchain_openai import OpenAIEmbeddings, ChatOpenAI
9
+ from langchain_community.vectorstores import Pinecone as PineconeVectorStore
10
  from pinecone import Pinecone
11
+ from supabase import create_client, Client
12
+ from langchain.memory import ConversationBufferWindowMemory
13
+ from langchain_core.messages import SystemMessage, HumanMessage
14
+ from langchain_core.tools import tool
15
+ from langgraph.graph import StateGraph, END
16
+ from langgraph.prebuilt import ToolNode, tools_condition
 
 
 
17
 
18
+ # Load environment variables
19
+ load_dotenv()
 
 
 
 
 
20
 
21
+ # Environment variables with better error handling
22
+ SUPABASE_URL = os.environ.get("SUPABASE_URL")
23
+ SUPABASE_KEY = os.environ.get("SUPABASE_KEY")
24
+ OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY")
25
+ LANGSMITH_API_KEY = os.environ.get("LANGSMITH_API_KEY")
26
+ PINECONE_API_KEY = os.environ.get("PINECONE_API_KEY")
27
+ PINECONE_INDEX = os.environ.get("PINECONE_INDEX", "natstudy")
28
 
29
+ # Check for required environment variables
30
+ required_vars = {
31
+ "SUPABASE_URL": SUPABASE_URL,
32
+ "SUPABASE_KEY": SUPABASE_KEY,
33
+ "OPENAI_API_KEY": OPENAI_API_KEY,
34
+ "PINECONE_API_KEY": PINECONE_API_KEY
35
+ }
36
+
37
+ missing_vars = [var for var, value in required_vars.items() if not value]
38
+ if missing_vars:
39
+ raise ValueError(f"Missing required environment variables: {', '.join(missing_vars)}")
40
+
41
+ # Model and vector store setup
42
+ try:
43
+ llm = ChatOpenAI(
44
+ model="gpt-4o-mini", # Use a more reliable model
45
+ temperature=0.0,
46
+ api_key=OPENAI_API_KEY
47
+ )
48
+
49
+ embeddings = OpenAIEmbeddings(
50
+ model="text-embedding-3-small", # Smaller model for cost efficiency
51
+ api_key=OPENAI_API_KEY
52
+ )
53
+
54
+ # Pinecone setup
55
+ pc = Pinecone(api_key=PINECONE_API_KEY)
56
+
57
+ vector_store = PineconeVectorStore(
58
+ embedding=embeddings,
59
+ index_name=PINECONE_INDEX,
60
+ pinecone_api_key=PINECONE_API_KEY
61
+ )
62
+
63
+ # Supabase
64
+ supabase: Client = create_client(SUPABASE_URL, SUPABASE_KEY)
65
+
66
+ except Exception as e:
67
+ print(f"Setup error: {e}")
68
+ # Create dummy implementations for development
69
+ llm = None
70
+ vector_store = None
71
+ supabase = None
72
 
73
  def log_interaction_cloud(user_id: str, question: str, answer: str):
74
+ if not supabase:
75
+ print("Supabase not configured")
76
+ return
77
+
78
  payload = {
79
  "user_id": user_id,
80
  "question": question,
 
83
  }
84
  try:
85
  res = supabase.table("logs").insert(payload).execute()
86
+ print("[Supabase Success]")
 
 
 
87
  except Exception as e:
88
+ print(f"[Supabase Exception]: {e}")
89
 
 
 
 
90
  def generate_user_id():
91
  return "user_" + uuid.uuid4().hex[:6]
92
 
93
+ # Initialize memory
94
  memory = ConversationBufferWindowMemory(
95
  memory_key="chat_history",
96
  return_messages=True,
97
  k=5
98
  )
99
 
 
 
 
100
  @tool(response_format="content_and_artifact")
101
  def retrieve(query: str):
102
+ """Retrieve information related to the query."""
103
+ if not vector_store:
104
+ return "Vector store not available", []
105
+
106
+ try:
107
+ retrieved_docs = vector_store.similarity_search(query, k=5)
108
+ serialized = "\n\n".join(
109
+ f"Source: {doc.metadata}\nContent: {doc.page_content}"
110
+ for doc in retrieved_docs
111
+ )
112
+ return serialized, retrieved_docs
113
+ except Exception as e:
114
+ return f"Retrieval error: {e}", []
115
 
116
  SYSTEM_MESSAGE_TEXT = (
117
+ "You are 'Einbürgerungs-Helfer', a highly skilled, friendly, and helpful administrative "
118
+ "assistant in Germany. Your only task is to support people who are eligible for naturalization "
119
+ "but have not yet applied. You provide accurate information about German citizenship requirements, "
120
+ "application processes, and related topics. You can respond in German, English, Turkish, or Arabic."
121
  )
122
 
123
+ class MyState(TypedDict):
124
+ messages: list
125
+ user_id: str
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
126
 
127
+ def query_or_respond(state: MyState):
128
+ if not llm:
129
+ return {"messages": state["messages"] + [{"role": "assistant", "content": "Service temporarily unavailable"}], "user_id": state.get("user_id", "unknown")}
130
+
131
+ try:
132
+ system_message = SystemMessage(SYSTEM_MESSAGE_TEXT)
133
+ raw_history = memory.load_memory_variables({})["chat_history"]
134
+
135
+ # Filter out repeated messages
136
+ filtered_history = [msg for msg in raw_history if not (
137
+ hasattr(msg, 'content') and msg.content and "I can only answer questions related" in msg.content
138
+ )]
139
+
140
+ prompt = [system_message] + filtered_history[-5:]
141
+ llm_with_tools = llm.bind_tools([retrieve])
142
+ response = llm_with_tools.invoke(prompt)
143
+
144
+ user_id = state.get("user_id", "unknown_user")
145
+ user_question = [m["content"] for m in state["messages"] if m.get("role") == "user"][-1]
146
+
147
+ memory.save_context({"input": user_question}, {"output": response.content})
148
+
149
+ if not getattr(response, "tool_calls", None):
150
+ log_interaction_cloud(user_id, user_question, response.content)
151
+
152
+ return {"messages": state["messages"] + [response], "user_id": user_id}
153
+ except Exception as e:
154
+ error_msg = f"Error processing request: {str(e)}"
155
+ return {"messages": state["messages"] + [{"role": "assistant", "content": error_msg}], "user_id": state.get("user_id", "unknown")}
156
+
157
+ def generate(state: MyState):
158
+ try:
159
+ last_message = state["messages"][-1]
160
+ user_id = state.get("user_id", "unknown_user")
161
+ user_question = [m["content"] for m in state["messages"] if m.get("role") == "user"][-1]
162
+ answer_text = getattr(last_message, "content", str(last_message))
163
+ log_interaction_cloud(user_id, user_question, answer_text)
164
+ return {"messages": state["messages"], "user_id": user_id}
165
+ except Exception as e:
166
+ print(f"Generate error: {e}")
167
+ return {"messages": state["messages"], "user_id": state.get("user_id", "unknown")}
168
+
169
+ # LangGraph setup
170
+ workflow = StateGraph(MyState)
171
+ workflow.add_node("query_or_respond", query_or_respond)
172
+ workflow.add_node("tool", ToolNode([retrieve]))
173
+ workflow.add_node("generate", generate)
174
+ workflow.add_conditional_edges("query_or_respond", tools_condition, {
175
+ "tool": "tool",
176
+ "generate": "generate"
177
+ })
178
+ workflow.add_edge("tool", "generate")
179
+ workflow.add_edge("generate", END)
180
+ workflow.set_entry_point("query_or_respond")
181
+ graph = workflow.compile()
182
+
183
+ # Gradio UI
184
  chat_histories = {}
185
 
186
+ def respond(user_message, chat_history, user_id):
187
+ if not user_message.strip():
188
+ return chat_history
189
+
190
+ if not user_id:
191
+ user_id = generate_user_id()
192
+
193
+ if user_id not in chat_histories:
194
+ chat_histories[user_id] = []
195
+
196
+ try:
197
+ state = {
198
+ "messages": [{"role": "user", "content": user_message}],
199
+ "user_id": user_id,
200
+ }
201
+
202
+ # Get response from graph
203
+ result = graph.invoke(state)
204
+ bot_response = result["messages"][-1].content if result["messages"] else "Sorry, I couldn't process your request."
205
+
206
+ new_history = chat_history + [[user_message, bot_response]]
207
+ chat_histories[user_id] = new_history
208
+
209
+ return new_history
210
+ except Exception as e:
211
+ error_response = f"I'm sorry, I encountered an error: {str(e)}"
212
+ return chat_history + [[user_message, error_response]]
213
 
214
  def clear_history():
215
  return "", []
216
 
217
+ # Multilingual description
218
+ description = """
219
+ ### Was ist **Einbürgerungs-Helfer**?
220
+ **Deutsch:** Einbürgerungs-Helfer hilft dir, den Weg zur deutschen Staatsangehörigkeit zu verstehen – einfach, mehrsprachig, datenschutzfreundlich.
221
+
222
+ **English:** Einbürgerungs-Helfer helps you understand how to get German citizenship – simple, multilingual, and privacy-friendly.
223
+
224
+ **Türkçe:** Einbürgerungs-Helfer, Almanya vatandaşlığına giden yolu anlamanıza yardımcı olur – basit, çok dilli ve gizliliğe saygılı.
225
+
226
+ **العربية:** يساعدك "Einbürgerungs-Helfer" على فهم كيفية الحصول على الجنسية الألمانية - بطريقة بسيطة، متعددة اللغات وتحترم الخصوصية.
227
+ """
228
+
229
+ # Create interface
230
+ with gr.Blocks(
231
+ theme=gr.themes.Soft(primary_hue="blue"),
232
+ css="""
233
+ @import url('https://fonts.googleapis.com/css2?family=Noto+Sans&display=swap');
234
+ * {
235
+ font-family: 'Noto Sans', sans-serif !important;
236
+ }
237
+ """
238
+ ) as iface:
239
+ gr.Markdown(f"# 🇩 Einbürgerungs-Helfer\n{description}")
240
 
 
 
241
  with gr.Row():
242
+ user_id_input = gr.Textbox(
243
+ label="Benutzer-ID (User ID)",
244
+ placeholder="z.B. user_123",
245
+ lines=1,
246
+ value=generate_user_id()
247
+ )
248
+ clear_btn = gr.Button("🗑️ Verlauf löschen")
249
+
250
+ chatbot = gr.Chatbot(
251
+ label="Chat",
252
+ show_copy_button=True,
253
+ value=[[None,
254
+ "Hallo! Ich bin hier, um Ihnen bei allen Fragen zur Einbürgerung in Deutschland zu helfen.\n\n"
255
+ "Sie können Ihre Frage auf Deutsch, Englisch, Türkisch oder Arabisch stellen.\n\n"
256
+ "Wie kann ich Ihnen heute helfen?"]]
257
+ )
258
+
259
+ msg_input = gr.Textbox(
260
+ placeholder="Frage eingeben...",
261
+ label="Deine Nachricht",
262
+ lines=1
263
+ )
264
+
265
+ submit_btn = gr.Button("Senden", variant="primary")
266
+
267
+ # Event handlers
268
+ submit_btn.click(
269
+ fn=respond,
270
+ inputs=[msg_input, chatbot, user_id_input],
271
+ outputs=[chatbot]
272
+ ).then(
273
+ lambda: "",
274
+ outputs=[msg_input]
275
+ )
276
+
277
+ msg_input.submit(
278
+ fn=respond,
279
+ inputs=[msg_input, chatbot, user_id_input],
280
+ outputs=[chatbot]
281
+ ).then(
282
+ lambda: "",
283
+ outputs=[msg_input]
284
+ )
285
+
286
+ clear_btn.click(
287
+ fn=clear_history,
288
+ outputs=[user_id_input, chatbot]
289
+ )
290
+
291
+ # Launch configuration for Hugging Face Spaces
292
+ if __name__ == "__main__":
293
+ iface.launch(
294
+ server_name="0.0.0.0",
295
+ server_port=7860,
296
+ share=False # Set to False for Spaces
297
+ )