Update langgraph_agent.py
Browse files- langgraph_agent.py +6 -85
langgraph_agent.py
CHANGED
@@ -5,8 +5,8 @@ import pandas as pd
|
|
5 |
from typing import Dict, List, Union
|
6 |
import re
|
7 |
|
8 |
-
from PIL import Image as PILImage
|
9 |
-
from huggingface_hub import InferenceClient
|
10 |
|
11 |
from langgraph.graph import START, StateGraph, MessagesState
|
12 |
from langgraph.prebuilt import tools_condition, ToolNode
|
@@ -73,8 +73,6 @@ def google_web_search(query: str) -> str:
|
|
73 |
return f"Error occurred while searching the web for '{query}'. Details: {str(e)}"
|
74 |
|
75 |
|
76 |
-
# HF_API_TOKEN is no longer directly needed for describe_image as that tool is removed.
|
77 |
-
# But keeping InferenceClient initialization for completeness if other HF tools might be added later.
|
78 |
HF_API_TOKEN = os.getenv("HF_API_TOKEN")
|
79 |
MODEL = os.getenv("MODEL")
|
80 |
HF_INFERENCE_CLIENT = None
|
@@ -133,50 +131,17 @@ def python_interpreter(code: str) -> Dict[str, str]:
|
|
133 |
except Exception as e:
|
134 |
return {"execution_error": str(e)}
|
135 |
|
136 |
-
# --- Youtube Tool (Remains the same) ---
|
137 |
-
@tool
|
138 |
-
def Youtube(url: str, question: str) -> Dict[str, str]:
|
139 |
-
"""
|
140 |
-
Tells about the YouTube video identified by the given URL, answering a question about it.
|
141 |
-
Note: This is a simulated response. In a real application, this would interact with a YouTube API
|
142 |
-
or a video analysis service to get actual video information and transcripts.
|
143 |
-
"""
|
144 |
-
print(f"Youtube called with URL: {url}, Question: {question}")
|
145 |
-
|
146 |
-
# Placeholder for actual YouTube API call.
|
147 |
-
# In a real scenario, you'd use a library like `google-api-python-client` for YouTube Data API
|
148 |
-
# or a dedicated video transcription/analysis service.
|
149 |
-
|
150 |
-
# Simulating the previous video content for demonstration
|
151 |
-
if "https://www.youtube.com/watch?v=1htKBjuUWec" in url or re.search(r'youtube\.com/watch\?v=|youtu\.be/', url):
|
152 |
-
return {
|
153 |
-
"video_url": url,
|
154 |
-
"question_asked": question,
|
155 |
-
"video_summary": "The video titled 'Teal'c coffee first time' shows a scene where several individuals are reacting to a beverage, presumably coffee, that Teal'c is trying for the first time. Key moments include: A person off-screen remarking, 'Wow this coffee's great'; another asking if it's 'cinnamon chicory tea oak'; and Teal'c reacting strongly to the taste or temperature, stating 'isn't that hot' indicating he finds it very warm.",
|
156 |
-
"details": {
|
157 |
-
"00:00:00": "Someone remarks, 'Wow this coffee's great I was just thinking that yeah is that cinnamon chicory tea oak'",
|
158 |
-
"00:00:11": "Teal'c takes a large gulp from a black mug",
|
159 |
-
"00:00:24": "Teal'c reacts strongly, someone asks 'isn't that hot'",
|
160 |
-
"00:00:26": "Someone agrees, 'extremely'"
|
161 |
-
}
|
162 |
-
}
|
163 |
-
else:
|
164 |
-
return {"error": "Invalid or unrecognized YouTube URL.", "url": url}
|
165 |
-
|
166 |
-
# --- END YOUTUBE TOOL ---
|
167 |
-
|
168 |
API_KEY = os.getenv("GEMINI_API_KEY")
|
169 |
-
HF_API_TOKEN = os.getenv("HF_SPACE_TOKEN")
|
170 |
GEMINI_API_KEY = os.getenv("GEMINI_API_KEY")
|
171 |
|
172 |
-
#
|
173 |
tools = [
|
174 |
multiply, add, subtract, divide, modulus,
|
175 |
wiki_search,
|
176 |
google_web_search,
|
177 |
read_file_content,
|
178 |
python_interpreter,
|
179 |
-
Youtube,
|
180 |
]
|
181 |
|
182 |
with open("prompt.txt", "r", encoding="utf-8") as f:
|
@@ -206,52 +171,8 @@ def build_graph(provider: str = "gemini"):
|
|
206 |
|
207 |
def assistant(state: MessagesState):
|
208 |
messages_to_send = [sys_msg] + state["messages"]
|
209 |
-
|
210 |
-
# --- IMPORTANT NOTE ON HANDLING BINARY BLOB DATA FOR MULTIMODAL LLMs ---
|
211 |
-
# When read_file_content returns a file_type of "image" or "audio",
|
212 |
-
# the agent should be able to send the actual binary data of that file
|
213 |
-
# as part of the message to the LLM. LangChain's ChatGoogleGenerativeAI
|
214 |
-
# supports this via content parts in HumanMessage.
|
215 |
-
#
|
216 |
-
# For this setup, we're assuming the framework (LangGraph/LangChain)
|
217 |
-
# will correctly handle passing the actual file content when read_file_content
|
218 |
-
# is called and its output indicates a media type.
|
219 |
-
#
|
220 |
-
# A more explicit implementation in the assistant node might look like this
|
221 |
-
# for real binary file handling if the framework doesn't do it implicitly:
|
222 |
-
#
|
223 |
-
# new_messages_to_send = []
|
224 |
-
# for msg in state["messages"]:
|
225 |
-
# if isinstance(msg, HumanMessage) and msg.tool_calls:
|
226 |
-
# # If a tool call to read_file_content happened in the previous turn
|
227 |
-
# # and it returned a media type, we might need to get the file data
|
228 |
-
# # and append it to the message parts. This logic is complex and
|
229 |
-
# # depends heavily on how tool outputs are structured and passed.
|
230 |
-
# # For simplicity in this template, we assume direct handling by the LLM
|
231 |
-
# # if the tool output indicates media, and the file itself is accessible
|
232 |
-
# # via the environment.
|
233 |
-
# pass # Keep original message, tool output will follow
|
234 |
-
# elif isinstance(msg, HumanMessage) and any(part.get("file_type") in ["image", "audio"] for part in msg.content if isinstance(part, dict)):
|
235 |
-
# # This is a conceptual example for if the HumanMessage itself contains file data
|
236 |
-
# # or a reference that needs to be resolved into data.
|
237 |
-
# # You'd need to load the actual file bytes here.
|
238 |
-
# # e.g., if msg.content was like: [{"type": "file_reference", "file_path": "image.png"}]
|
239 |
-
# # with open(msg.content[0]["file_path"], "rb") as f:
|
240 |
-
# # file_bytes = f.read()
|
241 |
-
# # new_messages_to_send.append(
|
242 |
-
# # HumanMessage(
|
243 |
-
# # content=[
|
244 |
-
# # {"type": "text", "text": "Here is the media content:"},
|
245 |
-
# # {"type": "image_data" if "image" in msg.content[0]["file_type"] else "audio_data", "data": base64.b64encode(file_bytes).decode('utf-8'), "media_type": "image/png" if "image" in msg.content[0]["file_type"] else "audio/mp3"}
|
246 |
-
# # ]
|
247 |
-
# # )
|
248 |
-
# # )
|
249 |
-
# else:
|
250 |
-
# new_messages_to_send.append(msg)
|
251 |
-
# llm_response = llm_with_tools.invoke([sys_msg] + new_messages_to_send)
|
252 |
-
# --- END IMPORTANT NOTE ---
|
253 |
|
254 |
-
llm_response = llm_with_tools.invoke(messages_to_send,{"recursion_limit": 25})
|
255 |
print(f"LLM Raw Response: {llm_response}")
|
256 |
return {"messages": [llm_response]}
|
257 |
|
@@ -265,4 +186,4 @@ def build_graph(provider: str = "gemini"):
|
|
265 |
return builder.compile()
|
266 |
|
267 |
if __name__ == "__main__":
|
268 |
-
pass
|
|
|
5 |
from typing import Dict, List, Union
|
6 |
import re
|
7 |
|
8 |
+
from PIL import Image as PILImage # Keep PIL for potential future use or if other parts depend on it, but describe_image is removed.
|
9 |
+
from huggingface_hub import InferenceClient # Keep InferenceClient for other potential HF uses, but describe_image is removed.
|
10 |
|
11 |
from langgraph.graph import START, StateGraph, MessagesState
|
12 |
from langgraph.prebuilt import tools_condition, ToolNode
|
|
|
73 |
return f"Error occurred while searching the web for '{query}'. Details: {str(e)}"
|
74 |
|
75 |
|
|
|
|
|
76 |
HF_API_TOKEN = os.getenv("HF_API_TOKEN")
|
77 |
MODEL = os.getenv("MODEL")
|
78 |
HF_INFERENCE_CLIENT = None
|
|
|
131 |
except Exception as e:
|
132 |
return {"execution_error": str(e)}
|
133 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
134 |
API_KEY = os.getenv("GEMINI_API_KEY")
|
135 |
+
HF_API_TOKEN = os.getenv("HF_SPACE_TOKEN") # Kept for potential future HF uses, but not for describe_image
|
136 |
GEMINI_API_KEY = os.getenv("GEMINI_API_KEY")
|
137 |
|
138 |
+
# Updated tools list without Youtube
|
139 |
tools = [
|
140 |
multiply, add, subtract, divide, modulus,
|
141 |
wiki_search,
|
142 |
google_web_search,
|
143 |
read_file_content,
|
144 |
python_interpreter,
|
|
|
145 |
]
|
146 |
|
147 |
with open("prompt.txt", "r", encoding="utf-8") as f:
|
|
|
171 |
|
172 |
def assistant(state: MessagesState):
|
173 |
messages_to_send = [sys_msg] + state["messages"]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
174 |
|
175 |
+
llm_response = llm_with_tools.invoke(messages_to_send, {"recursion_limit": 25})
|
176 |
print(f"LLM Raw Response: {llm_response}")
|
177 |
return {"messages": [llm_response]}
|
178 |
|
|
|
186 |
return builder.compile()
|
187 |
|
188 |
if __name__ == "__main__":
|
189 |
+
pass
|