mime

rootfs/agents/TextSummary/agent.py

@@ -0,0 +1,48 @@
+import copy
+
+from aios_kernel import CustomAIAgent, AgentMsg, AgentMsgType, AgentPrompt
+from aios_kernel.compute_task import ComputeTaskResultCode
+from knowledge.data.writer import split_text
+
+
+class TextSummaryAgent(CustomAIAgent):
+    def __init__(self):
+        super().__init__("TextSummary", "Text Summary", 128000)
+
+    async def _process_msg(self, msg: AgentMsg, workspace=None) -> AgentMsg:
+        if msg.msg_type is not AgentMsgType.TYPE_MSG:
+            return AgentMsg.create_error_resp(msg, "only support msg type")
+
+        if msg.body_mime is not None and msg.body_mime != "text/plain":
+            return AgentMsg.create_error_resp(msg, "only support text/plain mime type")
+
+        chunks = split_text(msg.body, separators=["\n\n", "\n"], chunk_size=4000, chunk_overlap=200, length_function=len)
+
+        prompt = AgentPrompt()
+        prompt.system_message = "Your job is to generate a summary based on the input."
+        if len(chunks) == 1:
+            prompt.append(AgentPrompt(chunks[0]))
+            resp = await self.do_llm_complection(prompt)
+            if resp.result_code != ComputeTaskResultCode.OK:
+                return msg.create_error_resp(resp.error_str)
+            return msg.create_resp_msg(resp.result_str)
+
+        segments = []
+        for i, chunk in enumerate(chunks):
+            seg_prompt = copy.deepcopy(prompt)
+            seg_prompt.append(AgentPrompt(chunk))
+            resp = await self.do_llm_complection(seg_prompt)
+            if resp.result_code != ComputeTaskResultCode.OK:
+                return msg.create_error_resp(resp.error_str)
+            segments.append(resp.result_str)
+
+        segments_str = "\n".join(segments)
+        prompt.append(AgentPrompt(f"以下文本分段之后的各段摘要，请合并生成一个完整摘要：\n{segments_str}"))
+        resp = await self.do_llm_complection(prompt)
+        if resp.result_code != ComputeTaskResultCode.OK:
+            return msg.create_error_resp(resp.error_str)
+        return msg.create_resp_msg(resp.result_str)
+
+
+def init():
+    return TextSummaryAgent()

rootfs/agents/Vision/agent.toml

@@ -0,0 +1,7 @@
+instance_id = "Vision"
+fullname = "Vision"
+llm_model_name = "gpt-4-vision-preview"
+
+[[prompt]]
+role = "system"
+content = """你的工作对用户输入的图片和视频做分析，并根据用户的意图做出回应。"""

src/aios/agent/agent.py

@@ -27,6 +27,7 @@ from ..environment.workspace_env import WorkspaceEnvironment
 from ..storage.storage import AIStorage
 
 from ..knowledge import *
+from . import video_utils, image_utils
 
 logger = logging.getLogger(__name__)
 
@@ -423,10 +424,38 @@ class AIAgent(BaseAIAgent):
     async def _create_openai_thread(self) -> str:
         return None
 
+    def check_and_to_base64(self, image_path: str) -> str:
+        if image_utils.is_file(image_path):
+            return image_utils.image_to_base64(image_path)
+        else:
+            return image_path
+
     async def _process_msg(self,msg:AgentMsg,workspace = None) -> AgentMsg:
         msg_prompt = AgentPrompt()
         if msg.msg_type == AgentMsgType.TYPE_GROUPMSG:
             need_process = False
+            if msg.is_image_msg():
+                image_prompt, images = msg.get_image_body()
+                if image_prompt is None:
+                    content = [[{"type": "text", "text": f"{msg.sender}'s message"}]]
+                    content.extend([{"type": "image_url", "url": self.check_and_to_base64(image)} for image in images])
+                    msg_prompt.messages = [{"role": "user", "content": content}]
+                else:
+                    content = [{"type": "text", "text": f"{msg.sender}:{image_prompt}"}]
+                    content.extend([{"type": "image_url", "url": self.check_and_to_base64(image)} for image in images])
+                    msg_prompt.messages = [{"role": "user", "content": content}]
+            elif msg.is_video_msg():
+                video_prompt, video = msg.get_video_body()
+                frames = video_utils.extract_frames(video)
+                if video_prompt is None:
+                    content = [{"type": "text", "text": f"{msg.sender}'s message"}]
+                    content.extend([{"type": "image_url", "url": frame} for frame in frames])
+                    msg_prompt.messages = [{"role": "user", "content": content}]
+                else:
+                    content = [{"type": "text", "text": f"{msg.sender}:{video_prompt}"}]
+                    content.extend([{"type": "image_url", "url": frame} for frame in frames])
+                    msg_prompt.messages = [{"role": "user", "content": content}]
+            else:
                 msg_prompt.messages = [{"role":"user","content":f"{msg.sender}:{msg.body}"}]
             session_topic = msg.target + "#" + msg.topic
             chatsession = AIChatSession.get_session(self.agent_id,session_topic,self.chat_db)
@@ -440,6 +469,24 @@ class AIAgent(BaseAIAgent):
                 chatsession.append(msg)
                 resp_msg = msg.create_group_resp_msg(self.agent_id,"")
                 return resp_msg
+        else:
+            if msg.is_image_msg():
+                image_prompt, images = msg.get_image_body()
+                if image_prompt is None:
+                    msg_prompt.messages = [{"role": "user", "content": [{"type": "image_url", "url": image} for image in images]}]
+                else:
+                    content = [{"type": "text", "text": image_prompt}]
+                    content.extend([{"type": "image_url", "url": image} for image in images])
+                    msg_prompt.messages = [{"role": "user", "content": content}]
+            elif msg.is_video_msg():
+                video_prompt, video = msg.get_video_body()
+                frames = video_utils.extract_frames(video)
+                if video_prompt is None:
+                    msg_prompt.messages = [{"role": "user", "content": [{"type": "image_url", "url": frame} for frame in frames]}]
+                else:
+                    content = [{"type": "text", "text": video_prompt}]
+                    content.extend([{"type": "image_url", "url": frame} for frame in frames])
+                    msg_prompt.messages = [{"role": "user", "content": content}]
             else:
                 msg_prompt.messages = [{"role":"user","content":msg.body}]
             session_topic = msg.get_sender() + "#" + msg.topic

src/aios/agent/agent_base.py

@@ -9,7 +9,7 @@ import time
 import re
 import shlex
 import json
-from typing import List
+from typing import List, Tuple
 
 from .ai_function import FunctionItem, AIFunction
 from ..proto.agent_msg import AgentMsg, AgentMsgType
@@ -410,6 +410,10 @@ class BaseAIAgent(abc.ABC):
     def get_max_token_size(self) -> int:
         pass
 
+    @abstractmethod
+    async def _process_msg(self,msg:AgentMsg,workspace = None) -> AgentMsg:
+        pass
+
     @classmethod
     def get_inner_functions(cls, env:Environment) -> (dict,int):
         if env is None:
@@ -445,10 +449,29 @@ class BaseAIAgent(abc.ABC):
         #logger.debug(f"Agent {self.agent_id} do llm token static system:{system_prompt_len},function:{function_token_len},history:{history_token_len},input:{input_len}, totoal prompt:{system_prompt_len + function_token_len + history_token_len} ")
         if inner_functions is None and env is not None:
             inner_functions,_ = BaseAIAgent.get_inner_functions(env)
+
+        model_name = self.get_llm_model_name()
+        if org_msg.is_video_msg() or org_msg.is_image_msg():
+            if model_name.startswith("gpt4"):
+                model_name = "gpt-4-vision-preview"
         if is_json_resp:
-            task_result:ComputeTaskResult = await ComputeKernel.get_instance().do_llm_completion(prompt,resp_mode="json",mode_name=self.get_llm_model_name(),max_token=self.get_max_token_size(),inner_functions=inner_functions,timeout=None)
+            task_result: ComputeTaskResult = await (ComputeKernel.get_instance()
+            .do_llm_completion(
+                prompt,
+                resp_mode="json",
+                mode_name=model_name,
+                max_token=self.get_max_token_size(),
+                inner_functions=inner_functions,
+                timeout=None))
         else:
-            task_result:ComputeTaskResult = await ComputeKernel.get_instance().do_llm_completion(prompt,resp_mode="text",mode_name=self.get_llm_model_name(),max_token=self.get_max_token_size(),inner_functions=inner_functions,timeout=None)
+            task_result: ComputeTaskResult = await (ComputeKernel.get_instance()
+            .do_llm_completion(
+                prompt,
+                resp_mode="text",
+                mode_name=model_name,
+                max_token=self.get_max_token_size(),
+                inner_functions=inner_functions,
+                timeout=None))
         if task_result.result_code != ComputeTaskResultCode.OK:
             logger.error(f"_do_llm_complection llm compute error:{task_result.error_str}")
             #error_resp = msg.create_error_resp(task_result.error_str)

src/aios/knowledge/data/writer.py

@@ -86,7 +86,7 @@ def _split_text_with_regex(
     return [s for s in splits if s != ""]
 
 
-def _split_text(
+def split_text(
         text: str,
         separators: List[str],
         chunk_size: int,
@@ -127,7 +127,7 @@ def _split_text(
             if not new_separators:
                 final_chunks.append(s)
             else:
-                other_info = _split_text(s, new_separators, chunk_size, chunk_overlap, length_function)
+                other_info = split_text(s, new_separators, chunk_size, chunk_overlap, length_function)
                 final_chunks.extend(other_info)
     if _good_splits:
         merged_text = _merge_splits(_good_splits, _separator, chunk_size, chunk_overlap, length_function)
@@ -197,7 +197,7 @@ class ChunkListWriter:
                 )
             )
 
-        text_list = _split_text(text, separators, chunk_size, chunk_overlap, length_function)
+        text_list = split_text(text, separators, chunk_size, chunk_overlap, length_function)
         chunk_list = []
         hash_obj = hashlib.sha256()
 

src/aios_kernel/image_utils.py

@@ -0,0 +1,22 @@
+import base64
+import os.path
+
+
+def to_base64(image_path: str) -> str:
+    """Convert image to base64."""
+    ext = os.path.splitext(image_path)[1]
+    with open(image_path, "rb") as image_file:
+        base64_image = base64.b64encode(image_file.read()).decode("utf-8")
+        return f"data:image/{ext};base64,{base64_image}"
+
+
+def is_file(image_path: str) -> bool:
+    return os.path.isfile(image_path)
+
+
+def is_base64(image_path: str) -> bool:
+    return image_path.startswith("data:image/")
+
+
+def is_url(image_path: str) -> bool:
+    return image_path.startswith("http://") or image_path.startswith("https://")

src/aios_kernel/video_utils.py

@@ -0,0 +1,18 @@
+import base64
+from typing import List
+
+import cv2
+
+
+def extract_frames(video_path: str) -> List[str]:
+    """Extract frames from video."""
+    frames = []
+    vidcap = cv2.VideoCapture(video_path)
+    while vidcap.isOpened():
+        success, image = vidcap.read()
+        if not success:
+            break
+        _, buffer = cv2.imencode(".jpg", image)
+        frames.append(base64.b64encode(buffer).decode("utf-8"))
+    vidcap.release()
+    return frames

src/component/agent_manager/agent_manager.py

@@ -130,14 +130,11 @@ class AgentManager:
                 logger.error(f"read agent.toml cfg from {agent_media} failed! unexpected error occurred: {str(e)}")
                 return None
 
-            agent_name = os.path.split(agent_media.full_path)[1]
-            spec = importlib.util.spec_from_file_location(agent_name, custom_agent)
-            the_api = importlib.util.module_from_spec(spec)
-            spec.loader.exec_module(the_api)
-            if not hasattr(the_api,"Agent"):
+            agent = runpy.run_path(custom_agent)
+            if "init" not in agent:
                 logger.error(f"read agent.toml cfg from {agent_media} failed! unexpected error occurred: {str(e)}")
                 return None
-            return the_api.Agent()
+            return agent["init"]()
 
 
 

src/component/openai_node/open_ai_node.py

@@ -11,6 +11,7 @@ import requests
 
 from aios import ComputeTask, ComputeTaskResult, ComputeTaskState, ComputeTaskType,ComputeTaskResultCode,ComputeNode,AIStorage,UserConfig
 
+
 logger = logging.getLogger(__name__)
 
 
@@ -92,15 +93,19 @@ class OpenAI_ComputeNode(ComputeNode):
     def _image_2_text(self, task: ComputeTask):
         logger.info('openai image_2_text')
         # 本地图片处理
-        def encode_image(image_path):
-            with open(image_path, "rb") as image_file:
-                return base64.b64encode(image_file.read()).decode('utf-8')
+
         headers = {
             "Content-Type": "application/json",
             "Authorization": f"Bearer {self.openai_api_key }"
         }
         model_name = task.params["model_name"]
-        base64_image = encode_image(task.params["image_path"])
+        image_path = task.params["image_path"]
+
+        if image_utils.is_file(image_path):
+            url = image_utils.to_base64(image_path)
+        else:
+            url = image_path
+
         payload = {
             "model": model_name,
             "messages": [
@@ -114,7 +119,7 @@ class OpenAI_ComputeNode(ComputeNode):
                         {
                             "type": "image_url",
                             "image_url": {
-                                "url": f"data:image/jpeg;base64,{base64_image}"
+                                "url": url
                             }
                         }
                     ]

src/requirements.txt

@@ -151,3 +151,5 @@ psycopg2-binary
 pyodbc
 oracledb
 html2text
+docx2txt
+opencv-python

src/service/aios_shell/aios_shell.py

@@ -238,12 +238,12 @@ class AIOS_Shell:
     def get_version(self) -> str:
         return "0.5.1"
 
-    async def send_msg(self,msg:str,target_id:str,topic:str,sender:str = None) -> str:
+    async def send_msg(self,msg:str,target_id:str,topic:str,sender:str = None, msg_mime:str=None) -> str:
         if sender == self.username:
             AIBus().get_default_bus().register_message_handler(self.username,self._user_process_msg)
 
         agent_msg = AgentMsg()
-        agent_msg.set(sender,target_id,msg)
+        agent_msg.set(sender,target_id,msg,body_mime=msg_mime)
         agent_msg.topic = topic
         resp = await AIBus.get_default_bus().send_message(agent_msg)
         if resp is not None: