Spaces:
Sleeping
Sleeping
File size: 32,770 Bytes
4001ecc |
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 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 |
from langgraph.graph import StateGraph, START, END
from typing_extensions import TypedDict, Annotated, Literal, Optional
from langchain_core.messages import AnyMessage, HumanMessage, SystemMessage
from langgraph.graph.message import add_messages
from langchain_mistralai import ChatMistralAI
from langchain_openai import ChatOpenAI
from langgraph.prebuilt import ToolNode, tools_condition
from langchain_core.runnables.graph import MermaidDrawMethod
from langchain_community.tools import DuckDuckGoSearchRun
from langchain_community.tools import WikipediaQueryRun
from langchain_community.utilities import WikipediaAPIWrapper
from langchain_aws import ChatBedrock
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_community.document_loaders import UnstructuredExcelLoader
# from langchain_google_vertexai import ChatVertexAI
# from langfuse.callback import CallbackHandler
import base64
import json
import time
import requests
# import boto3
from yt_dlp import YoutubeDL
import os
# from urllib.parse import urlparse, parse_qs
import re
from dotenv import load_dotenv
# Load env vars from .env file
load_dotenv()
# Initialize Langfuse CallbackHandler for LangGraph/Langchain (tracing)
# langfuse_handler = CallbackHandler()
######## STATE ########
class State(TypedDict):
"""
A class representing the state of the agent.
"""
question: str
messages: Annotated[list[AnyMessage], add_messages]
input_file: str
downloaded_file: Optional[str]
task_id: str
web_search_node_result: AnyMessage
thinking_node_result: AnyMessage
vision_node_result: AnyMessage
video_node_result: AnyMessage
audio_node_result: AnyMessage
code_node_result: AnyMessage
excel_node_result: AnyMessage
next_node: str
########################
######## MODELS ########
def get_general_model():
llm_provider = os.getenv("LLM_PROVIDER", "mistral")
if llm_provider == "mistral":
general_model = ChatMistralAI(
model="mistral-large-2411",#"ministral-8b-latest",#"mistral-small-latest",
temperature=0,
max_retries=2,
api_key=os.getenv("MISTRAL_API_KEY")
)
if llm_provider == "aws":
general_model = ChatBedrock(
model_id="arn:aws:bedrock:us-east-1:416545197702:inference-profile/us.amazon.nova-lite-v1:0",
# provider="amazon",
temperature=0,
region_name="eu-west-3",
aws_access_key_id=os.getenv("AWS_ACCESS_KEY_ID"),
aws_secret_access_key=os.getenv("AWS_SECRET_ACCESS_KEY")
)
return general_model
def get_big_model():
big_model = ChatMistralAI(
model="mistral-medium-2505",
temperature=0,
max_retries=2,
api_key=os.getenv("MISTRAL_API_KEY")
)
return big_model
def get_vision_model():
vlm_provider = os.getenv("VLM_PROVIDER", "mistral")
if vlm_provider == "openai":
print("Spawning Open AI VLM")
vision_model = ChatOpenAI(
model="gpt-4o",
temperature=0,
max_tokens=None,
timeout=None,
max_retries=2,
api_key=os.getenv("OPENAI_API_KEY"),
)
if vlm_provider == "mistral":
print("Spawning Mistral VLM")
vision_model = ChatMistralAI(
model="pixtral-12b-2409",#"mistral-small-latest","pixtral-large-latest",#
temperature=0,
max_retries=2,
api_key=os.getenv("MISTRAL_API_KEY")
)
return vision_model
def get_video_handler_model():
video_handler_model = ChatGoogleGenerativeAI(
model="gemini-2.0-flash",
temperature=0,
max_tokens=None,
timeout=None,
max_retries=2,
# other params...
)
return video_handler_model
def get_audio_handler_model():
audio_handler_model = ChatOpenAI(
model="gpt-4o-audio-preview-2024-12-17",#,gpt-4o-mini-audio-preview-2024-12-17",#
temperature=0,
max_tokens=None,
timeout=None,
max_retries=2,
api_key=os.getenv("OPENAI_API_KEY"),
)
return audio_handler_model
########################
######## Functions ########
def download_youtube_content(url: str, output_path: Optional[str] = None) -> None:
"""
Download YouTube content (single video or playlist) in MP4 format only.
Args:
url (str): URL of the YouTube video or playlist
output_path (str, optional): Directory to save the downloads. Defaults to './downloads'
"""
# Set default output path if none provided
if output_path is None:
output_path = os.path.join(os.getcwd(), 'downloads')
# Create output directory if it doesn't exist
os.makedirs(output_path, exist_ok=True)
# Configure yt-dlp options for MP4 only
ydl_opts = {
'format': 'bestvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best',
'merge_output_format': 'mp4',
'ignoreerrors': True,
'no_warnings': False,
'extract_flat': False,
# Disable all additional downloads
'writesubtitles': False,
'writethumbnail': False,
'writeautomaticsub': False,
'postprocessors': [{
'key': 'FFmpegVideoConvertor',
'preferedformat': 'mp4',
}],
# Clean up options
'keepvideo': False,
'clean_infojson': True
}
ydl_opts['outtmpl'] = os.path.join(output_path, '%(title)s.%(ext)s')
print("Detected single video URL. Downloading video...")
try:
with YoutubeDL(ydl_opts) as ydl:
# Download content
ydl.download([url])
print(f"\nDownload completed successfully! Files saved to: {output_path}")
except Exception as e:
print(f"An error occurred: {str(e)}")
result = os.listdir(output_path)
video_file_names = [x for x in result if re.match(r".*\.mp4$", x)]
if len(video_file_names) == 1:
video_file_name = video_file_names.pop()
video_file_name = f"{output_path}/{video_file_name}"
else:
video_file_name = None
for other_files in result:
if f"{output_path}/{other_files}" != video_file_name:
print(f"Removing file: {other_files}")
os.remove(os.path.join(output_path, other_files))
return video_file_name
web_search = DuckDuckGoSearchRun()
wikipedia_search = WikipediaQueryRun(api_wrapper=WikipediaAPIWrapper())
def download_input_file(task_id: str) -> str:
"""
Download the file specified in state input_file key.
You only need the task_id to download the file.
Args:
task_id (str): The task_id of the file to download.
Returns:
str: The path to the downloaded file.
"""
output_path = os.path.join(os.getcwd(), 'downloads')
api_url = os.getenv("DEFAULT_API_URL")
# Create output directory if it doesn't exist
os.makedirs(output_path, exist_ok=True)
# Construct the full URL
url = f"{api_url}/files/{task_id}"
try:
# Send a GET request to download the file
response = requests.get(url, stream=True)
response.raise_for_status() # Raise an error for bad status codes
headers = dict(response.headers)
attachement = headers["content-disposition"]
regex_result = re.search(r'filename="(.*)"', attachement)
filename = regex_result.group(1)
# Define the output file path
output_file_path = os.path.join(output_path, filename)
# Write the file to the output path
with open(output_file_path, 'wb') as file:
for chunk in response.iter_content(chunk_size=8192):
file.write(chunk)
print(f"File downloaded successfully and saved to: {output_file_path}")
return output_file_path
except requests.exceptions.RequestException as e:
print(f"An error occurred while downloading the file: {str(e)}")
return ""
########################
######## LLM associations ########
general_model = get_general_model()
big_model = get_big_model()
vision_model = get_vision_model()
video_handler_model = get_video_handler_model()
audio_handler_model = get_audio_handler_model()
########################
######## Nodes Definition ########
search_tools = [
web_search,
wikipedia_search,
]
download_file_tool = [ download_input_file ]
web_search_node_agent = general_model.bind_tools(search_tools, parallel_tool_calls=False)
def thinking_node(state: State) -> dict:
"""
A powerful node to answer general questions, reflection, maths, deduction, prediction.
This node does not handle files
This node does not handle images or pictures
This node does not handle videos
This node does not handle audio
This node does not handle code
Args:
state (State): A dictionary containing the current state of the agent, including the 'question' key which holds the question to be answered.
Returns:
dict: A dictionary containing the response from the web search node, with the key 'thinking_node_result' holding the list of messages generated by the general model.
"""
prompt = f"""
You are a powerful assistant that answers general questions, reflection, maths, deduction, prediction.
1. You need to fully understand the question
2. You must think hard about what is relevant in the question to make the best answer
3. If there are calculations or maths, you need to verify twice before answering.
4. Report your thought process in detail, explaining your reasoning step-by-step.
Here is the question {state['question']}
Now provide your response immediately without any preamble in text but not in markdown.
"""
state["thinking_node_result"] = state.get("thinking_node_result", "")
sys_msg = SystemMessage(content=prompt)
thinking_node_response = [general_model.invoke([sys_msg] + [state["thinking_node_result"]])]
thinking_node_response[-1].pretty_print()
return {
"thinking_node_result": thinking_node_response,
}
def code_node(state: State) -> dict:
"""
A powerful node to handle and understand code.
This node does not handle images or pictures
This node does not handle videos
This node does not handle audio
This node does not access the web
Args:
state (State): A dictionary containing the current state of the agent, including the 'question' key which holds the question to be answered.
Returns:
dict: A dictionary containing the response from the web search node, with the key 'code_node_result' holding the list of messages generated by the general model.
"""
with open(state["downloaded_file"], "r") as code_file:
code = code_file.read()
prompt = f"""
You are a powerful assistant that handle and understand code.
1. You need to fully understand the question.
2. You must think hard about the code and predict the result to answer the question.
3. Report your thought process in detail, explaining your reasoning step-by-step.
Here is the question : {state['question']}
Here is the code : {code}
Now provide your response immediately without any preamble in text but not in markdown.
"""
sys_msg = SystemMessage(content=prompt)
code_node_response = [general_model.invoke([sys_msg])]
code_node_response[-1].pretty_print()
return {
"code_node_result": code_node_response,
}
def web_search_node(state: State) -> dict:
"""
A powerful node to answer questions and make research on the web based on the question provided in the state.
This node does not handle files
This node does not handle images or pictures
This node does not handle videos
This node does not handle audio
This node does not handle code
Args:
state (State): A dictionary containing the current state of the agent, including the 'question' key which holds the question to be answered.
Returns:
dict: A dictionary containing the response from the web search node, with the key 'web_search_node_result' holding the list of messages generated by the general model.
"""
prompt = f"""
You are a powerful assistant that makes research on the web in order to give the best answer to the question.
1. You need to fully understand the question
2. You must think hard about what is relevant in the question to make the best search with write words
3. You must use the best of the tools you have to answer the question precisly
4. Report your thought process in detail, explaining your reasoning step-by-step.
5. You must not change the way words or identifiers are written in the web search results.
Here are the tools available:
web_search:
{web_search.description}
Args:
{web_search.args_schema}
Returns:
{web_search.response_format}
wikipedia_search:
{wikipedia_search.description}
Args:
{wikipedia_search.args_schema}
Returns:
{wikipedia_search.response_format}
Here is the question {state['question']}
Now provide your response immediately without any preamble in text but not in markdown.
"""
state["web_search_node_result"] = state.get("web_search_node_result", "")
sys_msg = SystemMessage(content=prompt)
web_search_node_response = [web_search_node_agent.invoke([sys_msg] + [state["web_search_node_result"]])]
web_search_node_response[-1].pretty_print()
return {
"web_search_node_result": web_search_node_response,
}
def vision_node(state: State) -> dict:
"""
Vision model that can analyze images and pictures and answer questions about them.
This node does not handle videos.
This node does not handle audio.
This node does not handle code.
Args:
state (State): A dictionary containing the current state of the agent, including the 'question' key which holds the question to be answered and the 'input_file' key which holds the path to the image file.
Returns:
dict: A dictionary containing the response from the vision node, with the key 'vision_node_result' holding the list of messages generated by the vision model.
"""
prompt = f"""
You are a powerful vision assistant, you can analyze images and answer question about the picture
1. You need to fully understand the question.
2. You must think hard about what is relevant in the image to make the best answer to the question.
3. Report your thought process in detail, explaining your reasoning step-by-step.
Here is the question {state['question']}
Now provide your response immediately without any preamble in text but not in markdown.
"""
image_base64 = ""
try:
with open(state["downloaded_file"], "rb") as image_file:
image_bytes = image_file.read()
image_base64 = base64.b64encode(image_bytes).decode("utf-8")
mistral_image_handling = {
"type": "image_url",
"image_url": f"data:image/png;base64,{image_base64}",
}
openai_image_handling = {
"type": "image",
"source_type": "base64",
"mime_type": "image/png", # or image/png, etc.
"data": image_base64,
}
vision_provider = os.getenv("VLM_PROVIDER", "mistral")
if vision_provider == "openai":
image_handling = openai_image_handling
else:
image_handling = mistral_image_handling
message = [
{
"role": "user",
"content": [
{
"type": "text",
"text": prompt,
},
image_handling
]
}
]
vision_node_response = [vision_model.invoke(
input=message,
# config={
# "callbacks": [langfuse_handler]
# }
)]
vision_node_response[-1].pretty_print()
return {
"vision_node_result": vision_node_response
}
except Exception as e:
# A butler should handle errors gracefully
error_msg = f"Error extracting text: {str(e)}"
print(error_msg)
return {}
def video_node(state: State) -> str:
"""
Video handler model that can analyze videos and answer questions about them.
This node does not handle images or pictures.
This node does not handle audio.
This node does not handle code.
Args:
state (State): A dictionary containing the current state of the agent, including the 'question' key which holds the question to be answered.
Returns:
dict: A dictionary containing the response from the video handler node, with the key 'video_node_result' holding the list of messages generated by the video handler model.
"""
prompt = f"""
You are a highly capable video analysis assistant. Your task is to watch and analyze the provided video content and answer the user's question as accurately and concisely as possible.
1. You need to fully understand the question.
2. Carefully observe the video, paying attention to relevant details, actions, and context.
3. Focus on the user's question.
4. If the question requires counting, identifying, or describing, be precise and clear in your response.
5. If you are unsure, state what you can infer from the video.
6. Do not make up information that is not visible or inferable from the video.
Here is the question {state['question']}
Now provide your response immediately without any preamble in text but not in markdown.
"""
if re.search(r'youtube\.com', state["question"]):
# More flexible regex pattern to match YouTube URLs
regex_result = re.search(r"(?P<youtube_url>https://(?:www\.)?youtube\.com/watch\?v=[a-zA-Z0-9_-]+)", state["question"])
if regex_result:
video_url = regex_result.group("youtube_url")
downloaded_video = download_youtube_content(url=video_url)
else:
# Fallback if regex doesn't match
print("Could not extract YouTube URL from question. Using question as fallback.")
downloaded_video = state["downloaded_file"]
else:
downloaded_video = state["downloaded_file"]
print(f"Downloaded video: {downloaded_video}")
video_mime_type = "video/mp4"
with open(downloaded_video, "rb") as video_file:
encoded_video = base64.b64encode(video_file.read()).decode("utf-8")
os.remove(downloaded_video)
message = [
{
"role": "user",
"content": [
{
"type": "text",
"text": prompt,
},
{
"type": "media",
"data": encoded_video, # Use base64 string directly
"mime_type": video_mime_type,
},
]
}
]
video_node_response = [video_handler_model.invoke(
input=message,
# config={
# "callbacks": [langfuse_handler]
# }
)]
video_node_response[-1].pretty_print()
return {
"video_node_result": video_node_response
}
def audio_node(state: State) -> str:
"""
Audio handler model that can analyze audio and answer questions about it.
This node does not handle images or pictures.
This node does not handle video.
This node does not handle code.
Args:
state (State): with question key inside
Returns:
dict: A dictionary containing the response from the video handler node, with the key 'audioo_node_result' holding the list of messages generated by the audio handler model.
"""
prompt = f"""
You are a highly capable audio analysis assistant. Your task is to listen to and analyze the provided audio content and answer the user's question as accurately and concisely as possible.
1. You need to fully understand the question.
2. Carefully listen to the audio, paying attention to relevant details, actions, and context.
3. Focus on the user's question.
4. If the question requires counting, identifying, or describing, be precise and clear in your response.
5. If you are unsure, state what you can infer from the audio.
6. Do not make up information that is not audible or inferable from the audio.
Here is the question {state['question']}
Now provide your response immediately without any preamble in text but not in markdown.
"""
downloaded_audio = state["downloaded_file"]
print(f"Downloaded audio: {downloaded_audio}")
audio_format = re.search(r'\.(\w+)$', downloaded_audio).group(1)
with open(downloaded_audio, "rb") as audio_file:
encoded_audio = base64.b64encode(audio_file.read()).decode()
os.remove(downloaded_audio)
message = [
{
"role": "user",
"content": [
{
"type": "text",
"text": prompt,
},
{
"type": "input_audio",
"input_audio": {
"data": encoded_audio,
"format": audio_format,
}
},
]
}
]
audio_node_response = [audio_handler_model.invoke(
input=message,
# config={
# "callbacks": [langfuse_handler]
# }
)]
audio_node_response[-1].pretty_print()
return {
"audio_node_result": audio_node_response
}
def excel_node(state: State):
"""
Excel handler model that can analyze excel files and answer questions about it.
This node does not handle images or pictures.
This node does not handle video.
This node does not handle code.
This node does not handle audio.
Args:
state (State): with question key inside
Returns:
dict: A dictionary containing the response from the excel handler node, with the key 'excel_node_result' holding the list of messages generated by the excel handler model.
"""
loader = UnstructuredExcelLoader(state["downloaded_file"], mode="elements")
docs = loader.load()
prompt = f"""
You are a powerful assistant which handles excel files.
1. You need to fully understand the question.
2. You must analyze the excel file to answer the question.
3. If the question requires counting, identifying, or describing, be precise and clear in your response.
4. Do not make up information that is not in the excel file.
Here is the question {state['question']}
Here is the excel file loaded in a Document object: {docs}. You will find htlm content of the file in the 'text_as_html' key.
Now provide your response immediately without any preamble in text but not in markdown.
"""
response = big_model.invoke(
input=prompt,
# config={
# "callbacks": [langfuse_handler]
# }
)
response.pretty_print()
return {
"excel_node_result": response
}
def format_answer_node(state: State):
"""
Format answer node that formats the answer of the last node.
This node does not handle images or pictures.
This node does not handle video.
This node does not handle audio.
This node does not handle code.
Args:
state (State): with question key inside, and all other nodes results
Returns:
dict: A dictionary containing the response from the format answer node, with the key 'format_answer_node_result' holding the list of messages generated by the format answer model.
"""
prompt = """
You are the best assistant for final answer formating.
1. You must not change the content of the response of the last node.
2. You must fully understand the question
3. You must return the answer by following hard the format and the constraints
4. Report your thought process in detail, explaining your reasoning step-by-step.
5. Conclude your answer with the following template:
FINAL ANSWER: [YOUR FINAL ANSWER]
## Response Format
- If asked for a number:
For exemple 'How many' or a question asking for a number result
- Provide the number without commas, dollar signs, percent signs, or any units (unless specified).
- Provide digits, not words
- If asked for a string:
- Write the string without articles (a, an, the).
- Don't answer a full sentence when a short version is enough.
- Do not use abbreviations (e.g., for cities).
- Write digits in text but (e.g., "one" instead of "1") unless specified otherwise.
- Start the first word with a capital letter.
- If asked for a comma-separated list:
- Apply the above rules for numbers and strings to each element in the list.
- And take care of having a space after each comma.
## Constraints
- You must not answer if the constraints above are not respected.
- Your final answer should be provided in the format: FINAL ANSWER: [YOUR FINAL ANSWER]
- Your final answer should be a number, a string, or a comma-separated list of numbers and/or strings, following the specified formatting rules.
Now provide your response immediately without any preamble in text but not in markdown.
"""
nodes_response = [HumanMessage(content="Here are the results of the previous nodes")]
question = [HumanMessage(content=state["question"])]
for node_result in ["web_search_node_result", "vision_node_result", "video_node_result", "audio_node_result", "thinking_node_result", "code_node_result", "excel_node_result"]:
result = state.get(node_result, "")
if result:
# Ensure result is a string. If it's a message object, extract its content.
if hasattr(result, "content"):
content = result.content
else:
content = str(result)
nodes_response.append(HumanMessage(content=content))
sys_msg = SystemMessage(content=prompt)
response = [general_model.invoke([sys_msg] + state["messages"]+ question + nodes_response)]
return {
"messages": response,
}
########################
######## Entry Node ########
def entry_node(state: State)-> str:
# System message
system_prompt = f"""
You are a powerful assistant that handle the user message and manage other nodes in order to provide the best answer to the question.
You do not handle images or pictures
You do not handle videos
You do not handle audio
You do not handle code
You do not handle excel files
1. You need to fully understand the subject of the question
2. You need to understand the subject of the question with the question itself and the file extension
For example of extensions:
- .py is for code
- .wav or .mp3 is for audio
- a youtube url is for video
- a .jpg, .png, .jpeg is for image
- a .xlsx or .xls is for excel
3. You must think hard about what is relevant in the question to make the best choice for the next node
4. You must not answer the question by yourself
5. Report your thought process in detail, explaining your reasoning step-by-step.
Here are the nodes you can choose:
- thinking_node: {thinking_node.__doc__}
- web_search_node: {web_search_node.__doc__}
- vision_node: {vision_node.__doc__}
- video_node: {video_node.__doc__}
- audio_node: {audio_node.__doc__}
- code_node: {code_node.__doc__}
- excel_node: {excel_node.__doc__}
Here is the question : {state['question']}
Here is the file : {state.get("input_file", "no file to handle")}
Now provide your response immediately.
You must always respect this format in lower case: next node <the node name you choose>.
"""
downloaded = ""
# If there's an input file, download it directly:
if state.get("input_file", None):
downloaded = download_input_file(state.get("task_id"))
sys_msg = SystemMessage(content=system_prompt)
entry_node_response = [general_model.invoke([sys_msg] + state["messages"])]
entry_node_response[-1].pretty_print()
regex_result = re.search(r'.*next.*(?P<next_node>thinking_node|web_search_node|vision_node|video_node|audio_node|code_node|excel_node)', entry_node_response[-1].content, re.IGNORECASE)
next_node = "END"
if regex_result:
# Extract the node name and remove any quotes around it
next_node = regex_result.group("next_node")
next_node = next_node.lower()
print(f"Next node to invoke: {next_node}")
return {
"next_node": next_node,
"downloaded_file": downloaded
}
########################
######## Build Graph ########
def buildweb_search_graph():
builder = StateGraph(State)
builder.add_node("web_search_node", web_search_node)
builder.add_node("tools", ToolNode(search_tools))
builder.add_edge(START, "web_search_node")
builder.add_conditional_edges(
"web_search_node",
tools_condition,
)
builder.add_edge("tools", "web_search_node")
builder.add_edge("web_search_node", END)
return builder.compile()
def build_graph():
builder = StateGraph(State)
builder.add_node("entry_node", entry_node)
builder.add_node("web_search_node", buildweb_search_graph())
builder.add_node("vision_node", vision_node)
builder.add_node("video_node", video_node)
builder.add_node("audio_node", audio_node)
builder.add_node("code_node", code_node)
builder.add_node("thinking_node", thinking_node)
builder.add_node("excel_node", excel_node)
builder.add_node("format_answer_node", format_answer_node)
builder.add_edge(START, "entry_node")
# Conditional routing from entry_node to specialized nodes
builder.add_conditional_edges(
"entry_node",
lambda state: state["next_node"],
{
"web_search_node": "web_search_node",
"vision_node": "vision_node",
"video_node": "video_node",
"audio_node": "audio_node",
"code_node": "code_node",
"excel_node": "excel_node",
"thinking_node": "thinking_node"
}
)
# After specialized node, go to END
builder.add_edge("web_search_node", "format_answer_node")
builder.add_edge("vision_node", "format_answer_node")
builder.add_edge("video_node", "format_answer_node")
builder.add_edge("audio_node", "format_answer_node")
builder.add_edge("code_node", "format_answer_node")
builder.add_edge("excel_node", "format_answer_node")
builder.add_edge("thinking_node", "format_answer_node")
builder.add_edge("format_answer_node", END)
return builder.compile()
########################
if __name__ == "__main__":
agent_graph = build_graph()
# Save the Mermaid diagram as text instead of trying to render as PNG
# This avoids issues with Pyppeteer browser launching
# with open("graph.png", "wb") as f:
# f.write(agent_graph.get_graph(xray=True).draw_mermaid_png())
# print("Graph saved as graph.png")
# print(vision_node.__doc__)
with open("./responses.json", "r") as responses:
json_responses = json.loads(responses.read())
# json_questions = [{
# "question": "The attached Excel file contains the sales of menu items for a local fast-food chain. What were the total sales that the chain made from food (not including drinks)? Express your answer in USD with two decimal places.",
# "file_name": "7bd855d8-463d-4ed5-93ca-5fe35145f733.xlsx",
# "task_id": "7bd855d8-463d-4ed5-93ca-5fe35145f733"
# }]
with open("questions.json", "r") as questions:
json_questions = json.loads(questions.read())
for input in json_questions:
question = input.get("question", "No question found")
file_name = input.get("file_name", "")
task_id = input.get("task_id", "")
print(f"QUESTION : {question}")
print(f"FILE: {file_name}")
user_prompt = [HumanMessage(content="Can you answer the question please ?")]
user_input = {"messages": user_prompt, "question": question, "input_file": file_name, "task_id": task_id}
messages = agent_graph.invoke(
input=user_input,
config={
"recursion_limit": 10,
# "callbacks": [langfuse_handler]
}
)
for m in messages['messages']:
m.pretty_print()
try:
regex_result = re.search(r"FINAL ANSWER:\s*(?P<answer>.*)$", messages['messages'][-1].content)
answer = regex_result.group("answer")
except:
regex_result = re.search(r"\s*(?P<answer>.*)$", messages['messages'][-1].content)
answer = regex_result.group("answer")
print(answer)
if answer == json_responses.get(task_id, ""):
print("The answer is correct !")
else:
print("The answer is incorrect !")
print(f"Expected: {json_responses.get(task_id, '')}")
print(f"Got: {answer}")
|