Spaces:
Sleeping
Sleeping
Update src/app/main_agent.py
Browse files- src/app/main_agent.py +41 -42
src/app/main_agent.py
CHANGED
@@ -1,50 +1,49 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
|
39 |
\"\"\"{transcript}\"\"\"
|
40 |
|
41 |
Now respond to the user's follow-up question: {user_question}
|
42 |
"""
|
43 |
-
|
44 |
-
|
45 |
|
46 |
-
|
47 |
-
|
48 |
-
return analysis_agent, follow_up_agent
|
49 |
|
|
|
50 |
|
|
|
1 |
+
from langchain_core.messages import BaseMessage, AIMessage
|
2 |
+
from langchain_core.runnables import RunnableLambda, Runnable
|
3 |
+
from langchain_community.llms import Ollama
|
4 |
+
from langchain.tools import Tool
|
5 |
+
from langgraph.graph import MessageGraph
|
6 |
+
import re
|
7 |
+
|
8 |
+
llm = Ollama(model="gemma3", temperature=0.0) # llama3.1
|
9 |
+
|
10 |
+
def create_agent(accent_tool_obj) -> tuple[Runnable, Runnable]:
|
11 |
+
accent_tool = Tool(
|
12 |
+
name="AccentAnalyzer",
|
13 |
+
func=accent_tool_obj.analyze,
|
14 |
+
description="Analyze a public MP4 video URL and determine the English accent with transcription."
|
15 |
+
)
|
16 |
+
|
17 |
+
def analyze_node(messages: list[BaseMessage]) -> AIMessage:
|
18 |
+
last_input = messages[-1].content
|
19 |
+
match = re.search(r'https?://\S+', last_input)
|
20 |
+
if match:
|
21 |
+
url = match.group()
|
22 |
+
result = accent_tool.func(url)
|
23 |
+
else:
|
24 |
+
result = "No valid video URL found in your message."
|
25 |
+
return AIMessage(content=result)
|
26 |
+
|
27 |
+
graph = MessageGraph()
|
28 |
+
graph.add_node("analyze_accent", RunnableLambda(analyze_node))
|
29 |
+
graph.set_entry_point("analyze_accent")
|
30 |
+
graph.set_finish_point("analyze_accent")
|
31 |
+
analysis_agent = graph.compile()
|
32 |
+
|
33 |
+
# Follow-up agent that uses transcript and responds to questions
|
34 |
+
def follow_up_node(messages: list[BaseMessage]) -> AIMessage:
|
35 |
+
user_question = messages[-1].content
|
36 |
+
transcript = accent_tool_obj.last_transcript or ""
|
37 |
+
prompt = f"""You are given this transcript of a video:
|
38 |
|
39 |
\"\"\"{transcript}\"\"\"
|
40 |
|
41 |
Now respond to the user's follow-up question: {user_question}
|
42 |
"""
|
43 |
+
response = llm.invoke(prompt)
|
44 |
+
return AIMessage(content=response)
|
45 |
|
46 |
+
follow_up_agent = RunnableLambda(follow_up_node)
|
|
|
|
|
47 |
|
48 |
+
return analysis_agent, follow_up_agent
|
49 |
|