semantio 0.0.7__py3-none-any.whl → 0.0.8__py3-none-any.whl

semantio/agent.py

@@ -22,19 +22,27 @@ from .memory import Memory
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
 
+
 class Agent(BaseModel):
-    # -*- Agent settings
+    """
+    An intelligent agent that combines LLM capabilities with dynamic knowledge base integration,
+    tool usage, and conversation memory. The agent can ingest external domain-specific content (via a dynamic document loader)
+    so that it answers queries based on that information.
+    """
     name: Optional[str] = Field(None, description="Name of the agent.")
     description: Optional[str] = Field(None, description="Description of the agent's role.")
     instructions: Optional[List[str]] = Field(None, description="List of instructions for the agent.")
-    model: Optional[str] = Field(None, description="This one is not in the use.")
+    model: Optional[str] = Field(None, description="This one is not in use.")
     show_tool_calls: bool = Field(False, description="Whether to show tool calls in the response.")
     markdown: bool = Field(False, description="Whether to format the response in markdown.")
     tools: Optional[List[BaseTool]] = Field(None, description="List of tools available to the agent.")
     user_name: Optional[str] = Field("User", description="Name of the user interacting with the agent.")
     emoji: Optional[str] = Field(":robot:", description="Emoji to represent the agent in the CLI.")
     rag: Optional[RAG] = Field(None, description="RAG instance for context retrieval.")
-    knowledge_base: Optional[Any] = Field(None, description="Knowledge base for domain-specific information.")
+    knowledge_base: Optional[Any] = Field(
+        None,
+        description="Domain-specific knowledge base content (e.g., loaded via a dynamic document loader)."
+    )
     llm: Optional[str] = Field(None, description="The LLM provider to use (e.g., 'groq', 'openai', 'anthropic').")
     llm_model: Optional[str] = Field(None, description="The specific model to use for the LLM provider.")
     llm_instance: Optional[BaseLLM] = Field(None, description="The LLM instance to use.")
@@ -57,120 +65,46 @@ class Agent(BaseModel):
         }
     )
     
-    # Allow arbitrary types
     model_config = ConfigDict(arbitrary_types_allowed=True)
 
     def __init__(self, **kwargs):
         super().__init__(**kwargs)
-        # Initialize the model and tools here if needed
+        # Initialize the LLM model and tools if needed.
         self._initialize_model()
-        # Initialize memory with config
+        # Initialize conversation memory with configuration.
         self.memory = Memory(
             max_context_length=self.memory_config.get("max_context_length", 4000),
             summarization_threshold=self.memory_config.get("summarization_threshold", 3000)
         )
-        # Initialize tools as an empty list if not provided
+        # Initialize tools as an empty list if not provided.
         if self.tools is None:
             self.tools = []
-        # Automatically discover and register tools if auto tool is enabled
+        # Automatically discover and register tools if auto_tool is enabled.
         if self.auto_tool and not self.tools:
             self.tools = self._discover_tools()
-        # Pass the LLM instance to each tool
+        # Pass the LLM instance to each tool.
         for tool in self.tools:
             tool.llm = self.llm_instance
-        # Initialize the SentenceTransformer model for semantic matching
+        # Initialize the SentenceTransformer model for semantic matching.
         self.semantic_model = SentenceTransformer('all-MiniLM-L6-v2')
-        # Initialize RAG if not provided
+        # Initialize default RAG if not provided.
         if self.rag is None:
             self.rag = self._initialize_default_rag()
-        # Automatically generate API if api=True
+        # Automatically generate API if api=True.
         if self.api:
             self._generate_api()
 
-
-    def _generate_response_from_image(self,message: str, image: Union[str, Image], markdown: bool = False, **kwargs) -> str:
-        """
-        Send the image to the LLM for analysis if the LLM supports vision.
-        Supports both local images (PIL.Image) and image URLs.
-        """
-        try:
-            # Check if the LLM supports vision
-            if not self.llm_instance or not self.llm_instance.supports_vision:
-                raise ValueError("Vision is not supported for the current model.")
-            prompt = self._build_prompt(message, context=None)
-            # Handle image URL
-            if isinstance(image, str) and image.startswith("http"):
-                # Directly pass the URL to the LLM
-                return self.llm_instance.generate_from_image_url(prompt,image, **kwargs)
-
-            # Handle local image (PIL.Image)
-            elif isinstance(image, Image):
-                # Convert the image to bytes
-                if image.mode == "RGBA":
-                    image = image.convert("RGB")  # Convert RGBA to RGB
-                image_bytes = io.BytesIO()
-                image.save(image_bytes, format="JPEG")  # Save as PNG (or any supported format)
-                image_bytes = image_bytes.getvalue()
-
-                # Generate response using base64-encoded image bytes
-                return self.llm_instance.generate_from_image(prompt,image_bytes, **kwargs)
-
-            else:
-                raise ValueError("Unsupported image type. Provide either a URL or a PIL.Image.")
-
-        except Exception as e:
-            logger.error(f"Failed to generate response from image: {e}")
-            return f"An error occurred while processing the image: {e}"
-        
-    def _discover_tools(self) -> List[BaseTool]:
-        """
-        Automatically discover and register tools from the 'tools' directory.
-        """
-        tools = []
-        tools_dir = Path(__file__).parent / "tools"
-
-        if not tools_dir.exists():
-            logger.warning(f"Tools directory not found: {tools_dir}")
-            return tools
-
-        # Iterate over all Python files in the 'tools' directory
-        for file in tools_dir.glob("*.py"):
-            if file.name == "base_tool.py":
-                continue  # Skip the base tool file
-
-            try:
-                # Import the module
-                module_name = file.stem
-                module = importlib.import_module(f"semantio.tools.{module_name}")
-
-                # Find all classes that inherit from BaseTool
-                for name, obj in module.__dict__.items():
-                    if isinstance(obj, type) and issubclass(obj, BaseTool) and obj != BaseTool:
-                        # Instantiate the tool and add it to the list
-                        tools.append(obj())
-                        logger.info(f"Registered tool: {obj.__name__}")
-            except Exception as e:
-                logger.error(f"Failed to load tool from {file}: {e}")
-
-        return tools
-    
-    def _get_tool_descriptions(self) -> str:
-        """Generate a description of all available tools for the LLM prompt."""
-        return "\n".join(
-            f"{tool.name}: {tool.description}" for tool in self.tools
-        )
-    
     def _initialize_model(self):
-        """Initialize the model based on the provided configuration."""
+        """Initialize the LLM model based on the provided configuration."""
         if self.llm_instance is not None:
-            return  # LLM is already initialized, do nothing
+            return  # Already initialized.
         if self.llm is None:
             raise ValueError("llm must be specified.")
     
-        # Get the API key from the environment or the provided configuration
+        # Retrieve API key from configuration or environment variable.
         api_key = getattr(self, 'api_key', None) or os.getenv(f"{self.llm.upper()}_API_KEY")
     
-        # Map LLM providers to their respective classes and default models
+        # Map LLM providers to their respective classes and default models.
         llm_providers = {
             "groq": {
                 "class": "GroqLlm",
@@ -198,28 +132,23 @@ class Agent(BaseModel):
             },
         }
 
-        # Normalize the LLM provider name (case-insensitive)
         llm_provider = self.llm.lower()
-
         if llm_provider not in llm_providers:
-            raise ValueError(f"Unsupported LLM provider: {self.llm}. Supported providers are: {list(llm_providers.keys())}")
+            raise ValueError(f"Unsupported LLM provider: {self.llm}. Supported providers: {list(llm_providers.keys())}")
 
-        # Get the LLM class and default model
         llm_config = llm_providers[llm_provider]
         llm_class_name = llm_config["class"]
         default_model = llm_config["default_model"]
-
-        # Use the user-provided model or fallback to the default model
         model_to_use = self.llm_model or default_model
 
-        # Dynamically import and initialize the LLM class
+        # Dynamically import and initialize the LLM class.
         module_name = f"semantio.llm.{llm_provider}"
         llm_module = importlib.import_module(module_name)
         llm_class = getattr(llm_module, llm_class_name)
         self.llm_instance = llm_class(model=model_to_use, api_key=api_key)
 
     def _initialize_default_rag(self) -> RAG:
-        """Initialize a default RAG instance with a dummy vector store."""
+        """Initialize a default RAG instance using a dummy vector store."""
         vector_store = VectorStore()
         retriever = Retriever(vector_store)
         return RAG(retriever)
@@ -227,45 +156,129 @@ class Agent(BaseModel):
     def print_response(
         self,
         message: Optional[Union[str, Image, List, Dict]] = None,
+        image: Optional[Union[str, Image]] = None,
         stream: bool = False,
         markdown: bool = False,
         team: Optional[List['Agent']] = None,
         **kwargs,
     ) -> Union[str, Dict]:
-        """Print the agent's response to the console and return it."""
-    
-        # Store user message if provided
+        """
+        Generate and print the agent's response while storing conversation history.
+        If an image is provided (either via the 'image' parameter or if 'message' is a PIL.Image),
+        the agent processes it accordingly.
+        If a team is provided (or if self.team is set), only the aggregated final response is returned.
+        """
+        # Handle image input first.
+        if image is not None:
+            response = self._generate_response_from_image(message or "", image, markdown=markdown, **kwargs)
+            print(response)
+            if response:
+                self.memory.add_message(role="agent", content=response)
+            return response
+
+        if isinstance(message, Image):
+            response = self._generate_response_from_image("", message, markdown=markdown, **kwargs)
+            print(response)
+            if response:
+                self.memory.add_message(role="agent", content=response)
+            return response
+
+        # For text input, add the user message to memory.
         if message and isinstance(message, str):
             self.memory.add_message(role="user", content=message)
 
+        # If a team is provided (or if self.team exists), generate an aggregated final response.
+        if team is None and self.team is not None:
+            team = self.team
+
+        if team is not None:
+            # Instead of printing individual team outputs, call each agent's _generate_response
+            # to capture their outputs silently.
+            aggregated_responses = []
+            for agent in team:
+                resp = agent._generate_response(message, markdown=markdown, **kwargs)
+                aggregated_responses.append(f"**{agent.name}:**\n\n{resp}")
+            final_response = "\n\n".join(aggregated_responses)
+            print(final_response)
+            self.memory.add_message(role="agent", content=final_response)
+            return final_response
+
+        # Standard text response processing.
         if stream:
-            # Handle streaming response
             response = ""
             for chunk in self._stream_response(message, markdown=markdown, **kwargs):
                 print(chunk, end="", flush=True)
                 response += chunk
-            # Store agent response
             if response:
-                self.memory.add_message(role="assistant", content=response)
-            print()  # New line after streaming
+                self.memory.add_message(role="agent", content=response)
+            print()
             return response
         else:
-            # Generate and return the response
-            response = self._generate_response(message, markdown=markdown, team=team, **kwargs)
-            print(response)  # Print the response to the console
-            # Store agent response
+            response = self._generate_response(message, markdown=markdown, **kwargs)
+            print(response)
             if response:
-                self.memory.add_message(role="assistant", content=response)
+                self.memory.add_message(role="agent", content=response)
             return response
 
-
     def _stream_response(self, message: str, markdown: bool = False, **kwargs) -> Iterator[str]:
-        """Stream the agent's response."""
-        # Simulate streaming by yielding chunks of the response
+        """Simulate streaming of the agent's response."""
         response = self._generate_response(message, markdown=markdown, **kwargs)
         for chunk in response.split():
             yield chunk + " "
 
+    def _generate_response_from_image(self, message: str, image: Union[str, Image], markdown: bool = False, **kwargs) -> str:
+        """
+        Process an image by sending it to the LLM for analysis if the LLM supports vision.
+        Supports both image URLs and local PIL.Image objects.
+        """
+        try:
+            if not self.llm_instance or not getattr(self.llm_instance, "supports_vision", False):
+                raise ValueError("Vision is not supported for the current model.")
+            prompt = self._build_prompt(message, context=None)
+            if isinstance(image, str) and image.startswith("http"):
+                return self.llm_instance.generate_from_image_url(prompt, image, **kwargs)
+            elif isinstance(image, Image):
+                if image.mode == "RGBA":
+                    image = image.convert("RGB")
+                image_bytes = io.BytesIO()
+                image.save(image_bytes, format="JPEG")
+                image_bytes = image_bytes.getvalue()
+                return self.llm_instance.generate_from_image(prompt, image_bytes, **kwargs)
+            else:
+                raise ValueError("Unsupported image type. Provide either a URL or a PIL.Image.")
+        except Exception as e:
+            logger.error(f"Failed to generate response from image: {e}")
+            return f"An error occurred while processing the image: {e}"
+
+    def _discover_tools(self) -> List[BaseTool]:
+        """
+        Automatically discover and register tools from the 'tools' directory.
+        """
+        tools = []
+        tools_dir = Path(__file__).parent / "tools"
+        if not tools_dir.exists():
+            logger.warning(f"Tools directory not found: {tools_dir}")
+            return tools
+        for file in tools_dir.glob("*.py"):
+            if file.name == "base_tool.py":
+                continue  # Skip the base tool file.
+            try:
+                module_name = file.stem
+                module = importlib.import_module(f"semantio.tools.{module_name}")
+                for name, obj in module.__dict__.items():
+                    if isinstance(obj, type) and issubclass(obj, BaseTool) and obj != BaseTool:
+                        tools.append(obj())
+                        logger.info(f"Registered tool: {obj.__name__}")
+            except Exception as e:
+                logger.error(f"Failed to load tool from {file}: {e}")
+        return tools
+
+    def _get_tool_descriptions(self) -> str:
+        """
+        Generate a description of all available tools for inclusion in the LLM prompt.
+        """
+        return "\n".join(f"{tool.name}: {tool.description}" for tool in self.tools)
+
     def register_tool(self, tool: BaseTool):
         """Register a tool for the agent."""
         if self.tools is None:
@@ -274,10 +287,9 @@ class Agent(BaseModel):
     
     def _analyze_query_and_select_tools(self, query: str) -> List[Dict[str, Any]]:
         """
-        Use the LLM to analyze the query and dynamically select tools.
-        Returns a list of tool calls, each with the tool name and input.
+        Use the LLM to analyze the query and dynamically select the most appropriate tools.
+        Returns a list of tool calls (tool name and input).
         """
-        # Create a prompt for the LLM to analyze the query and select tools
         prompt = f"""
         You are an AI agent that helps analyze user queries and select the most appropriate tools.
         Below is a list of available tools and their functionalities:
@@ -301,211 +313,191 @@ class Agent(BaseModel):
             }}
         ]
         """
-
         try:
-            # Call the LLM to generate the response
             response = self.llm_instance.generate(prompt=prompt)
-            # Parse the response as JSON
             tool_calls = json.loads(response)
             return tool_calls
         except Exception as e:
             logger.error(f"Failed to analyze query and select tools: {e}")
             return []
     
-
     def _generate_response(self, message: str, markdown: bool = False, team: Optional[List['Agent']] = None, **kwargs) -> str:
         """Generate the agent's response, including tool execution and context retrieval."""
-        # Use the specified team if provided
         if team is not None:
             return self._generate_team_response(message, team, markdown=markdown, **kwargs)
-        # Initialize tool_outputs as an empty dictionary
+        
         tool_outputs = {}
         responses = []
         tool_calls = []
-        # Use the LLM to analyze the query and dynamically select tools when auto_tool is enabled
+
         if self.auto_tool:
             tool_calls = self._analyze_query_and_select_tools(message)
         else:
-            # Check if tools are provided
             if self.tools:
                 tool_calls = [
                     {
                         "tool": tool.name,
-                        "input": {
-                            "query": message,  # Use the message as the query
-                            "context": None,  # No context provided by default
-                        }
+                        "input": {"query": message, "context": None}
                     }
                     for tool in self.tools
                 ]
 
-        # Execute tools if any are detected
         if tool_calls:
             for tool_call in tool_calls:
                 tool_name = tool_call["tool"]
                 tool_input = tool_call["input"]
-
-                # Find the tool
                 tool = next((t for t in self.tools if t.name.lower() == tool_name.lower()), None)
                 if tool:
                     try:
-                        # Execute the tool
                         tool_output = tool.execute(tool_input)
-                        response = f"Tool '{tool_name}' executed. Output: {tool_output}"
+                        response_text = f"Tool '{tool_name}' executed. Output: {tool_output}"
                         if self.show_tool_calls:
-                            response = f"**Tool Called:** {tool_name}\n\n{response}"
-                        responses.append(response)
-
-                        # Store the tool output for collaboration
+                            response_text = f"**Tool Called:** {tool_name}\n\n{response_text}"
+                        responses.append(response_text)
                         tool_outputs[tool_name] = tool_output
                     except Exception as e:
-                        logger.error(f"Tool called:** {tool_name}\n\n{response}")
+                        logger.error(f"Error executing tool '{tool_name}': {e}")
                         responses.append(f"An error occurred while executing the tool '{tool_name}': {e}")
                 else:
                     responses.append(f"Tool '{tool_name}' not found.")
 
-        # If multiple tools were executed, combine their outputs for analysis
         if tool_outputs:
             try:
-                # Prepare the context for the LLM
                 context = {
                     "conversation_history": self.memory.get_context(self.llm_instance),
                     "tool_outputs": tool_outputs,
                     "rag_context": self.rag.retrieve(message) if self.rag else None,
                     "knowledge_base": self._get_knowledge_context(message) if self.knowledge_base else None,
                 }
-                # 3. Build a memory-aware prompt.
                 prompt = self._build_memory_prompt(message, context)
-                # To (convert MemoryEntry objects to dicts and remove metadata):
                 memory_entries = [{"role": e.role, "content": e.content} for e in self.memory.storage.retrieve()]
-                # Generate a response using the LLM
                 llm_response = self.llm_instance.generate(prompt=prompt, context=context, memory=memory_entries, **kwargs)
                 responses.append(f"**Analysis:**\n\n{llm_response}")
             except Exception as e:
                 logger.error(f"Failed to generate LLM response: {e}")
                 responses.append(f"An error occurred while generating the analysis: {e}")
-        if not self.tools and not tool_calls:
-            # If no tools were executed, proceed with the original logic
-            # Retrieve relevant context using RAG
-            rag_context = self.rag.retrieve(message) if self.rag else None
-            # Retrieve relevant context from the knowledge base (API result)
-            # knowledge_base_context = None
-            # if self.knowledge_base:
-            #     # Flatten the knowledge base
-            #     flattened_data = self._flatten_data(self.knowledge_base)
-            #     # Find all relevant key-value pairs in the knowledge base
-            #     relevant_values = self._find_all_relevant_keys(message, flattened_data)
-            #     if relevant_values:
-            #         knowledge_base_context = ", ".join(relevant_values)
-
-            # Combine both contexts (RAG and knowledge base)
+        elif not self.tools and not tool_calls:
             context = {
                 "conversation_history": self.memory.get_context(self.llm_instance),
-                "rag_context": rag_context,
+                "rag_context": self.rag.retrieve(message) if self.rag else None,
                 "knowledge_base": self._get_knowledge_context(message),
             }
-            # Prepare the prompt with instructions, description, and context
-            # 3. Build a memory-aware prompt.
             prompt = self._build_memory_prompt(message, context)
-            # To (convert MemoryEntry objects to dicts and remove metadata):
             memory_entries = [{"role": e.role, "content": e.content} for e in self.memory.storage.retrieve()]
-
-            # Generate the response using the LLM
             response = self.llm_instance.generate(prompt=prompt, context=context, memory=memory_entries, **kwargs)
-
-
-            # Format the response based on the json_output flag
             if self.json_output:
                 response = self._format_response_as_json(response)
-
-            # Validate the response against the expected_output
             if self.expected_output:
                 response = self._validate_response(response)
-
             if markdown:
                 return f"**Response:**\n\n{response}"
             return response
         return "\n\n".join(responses)
     
-    # Modified prompt construction with memory integration
+    def _generate_team_response(self, message: str, team: List['Agent'], markdown: bool = False, **kwargs) -> str:
+        """
+        Generate a final aggregated response using a team of assistants.
+        This method calls each team member's internal _generate_response (without printing)
+        and aggregates the results into a single output.
+        """
+        team_responses = []
+        for agent in team:
+            resp = agent._generate_response(message, markdown=markdown, **kwargs)
+            team_responses.append(f"**{agent.name}:**\n\n{resp}")
+        return "\n\n".join(team_responses)
+
     def _build_memory_prompt(self, user_input: str, context: dict) -> str:
-        """Enhanced prompt builder with memory context."""
+        """Construct a prompt that incorporates role, instructions, conversation history, and external context."""
         prompt_parts = []
-        
         if self.description:
             prompt_parts.append(f"# ROLE\n{self.description}")
-            
         if self.instructions:
-            prompt_parts.append(f"# INSTRUCTIONS\n" + "\n".join(f"- {i}" for i in self.instructions))
-            
-        if context['conversation_history']:
+            prompt_parts.append("# INSTRUCTIONS\n" + "\n".join(f"- {i}" for i in self.instructions))
+        if context.get('conversation_history'):
             prompt_parts.append(f"# CONVERSATION HISTORY\n{context['conversation_history']}")
-            
-        if context['knowledge_base']:
+        if context.get('knowledge_base'):
             prompt_parts.append(f"# KNOWLEDGE BASE\n{context['knowledge_base']}")
-            
         prompt_parts.append(f"# USER INPUT\n{user_input}")
-        
         return "\n\n".join(prompt_parts)
         
+    def _summarize_text(self, text: str) -> str:
+        """
+        Summarize the provided text using the LLM.
+        Adjust the prompt as needed.
+        """
+        prompt = f"Summarize the following text concisely:\n\n{text}\n\nSummary:"
+        summary = self.llm_instance.generate(prompt=prompt)
+        return summary.strip()
+
     def _get_knowledge_context(self, message: str) -> str:
-        """Retrieve and format knowledge base context."""
+        """
+        Retrieve context from the knowledge base.
+        For JSON documents, use the "flattened" field.
+        For other documents (e.g., website, YouTube) use the "text" field.
+        If the combined text is too long, break it into chunks and summarize each chunk.
+        """
         if not self.knowledge_base:
             return ""
+        texts = []
+        for doc in self.knowledge_base:
+            if isinstance(doc, dict):
+                if "flattened" in doc:
+                    # Join all values from the flattened key/value pairs.
+                    flattened_text = " ".join(str(v) for item in doc["flattened"] for v in item.values())
+                    texts.append(flattened_text)
+                elif "text" in doc:
+                    texts.append(doc["text"])
+                else:
+                    texts.append(" ".join(str(v) for v in doc.values()))
+            else:
+                texts.append(str(doc))
+        combined_text = "\n".join(texts)
         
-        flattened = self._flatten_data(self.knowledge_base)
-        relevant = self._find_all_relevant_keys(message, flattened)
-        return "\n".join(f"- {item}" for item in relevant) if relevant else ""
-    def _generate_team_response(self, message: str, team: List['Agent'], markdown: bool = False, **kwargs) -> str:
-        """Generate a response using a team of assistants."""
-        responses = []
-        for agent in team:
-            response = agent.print_response(message, markdown=markdown, **kwargs)
-            responses.append(f"**{agent.name}:**\n\n{response}")
-        return "\n\n".join(responses)
+        # If the combined text is very long, break it into chunks and summarize.
+        max_words = 1000
+        words = combined_text.split()
+        if len(words) > max_words:
+            chunks = []
+            for i in range(0, len(words), max_words):
+                chunk = " ".join(words[i:i+max_words])
+                chunks.append(chunk)
+            # Summarize each chunk.
+            summaries = [self._summarize_text(chunk) for chunk in chunks]
+            final_context = "\n".join(summaries)
+            return final_context
+        else:
+            return combined_text
+
 
+
+
+    
     def _build_prompt(self, message: str, context: Optional[List[Dict]]) -> str:
-        """Build the prompt using instructions, description, and context."""
+        """Build a basic prompt including description, instructions, context, and user input."""
         prompt_parts = []
-
-        # Add description if available
         if self.description:
             prompt_parts.append(f"Description: {self.description}")
-
-        # Add instructions if available
         if self.instructions:
-            instructions = "\n".join(self.instructions)
-            prompt_parts.append(f"Instructions: {instructions}")
-
-        # Add context if available
+            prompt_parts.append("Instructions: " + "\n".join(self.instructions))
         if context:
             prompt_parts.append(f"Context: {context}")
-
-        # Add the user's message
         prompt_parts.append(f"User Input: {message}")
-
         return "\n\n".join(prompt_parts)
 
     def _format_response_as_json(self, response: str) -> Union[Dict, str]:
-        """Format the response as JSON if json_output is True."""
+        """Attempt to extract and format a JSON response."""
         try:
-            # Use regex to extract JSON from the response (e.g., within ```json ``` blocks)
             json_match = re.search(r'```json\s*({.*?})\s*```', response, re.DOTALL)
             if json_match:
-                # Extract the JSON part and parse it
                 json_str = json_match.group(1)
-                return json.loads(json_str)  # Return the parsed JSON object (a dictionary)
+                return json.loads(json_str)
             else:
-                # If no JSON block is found, try to parse the entire response as JSON
-                return json.loads(response)  # Return the parsed JSON object (a dictionary)
+                return json.loads(response)
         except json.JSONDecodeError:
-            # If the response is not valid JSON, wrap it in a dictionary
-            return {"response": response}  # Return a dictionary with the response as a string
+            return {"response": response}
 
     def normalize_key(self, key: str) -> str:
-        """
-        Normalize a key by converting it to lowercase and replacing spaces with underscores.
-        """
+        """Normalize a key by converting to lowercase and replacing spaces with underscores."""
         return key.lower().replace(" ", "_")
     
     def match_key(self, expected_key, response_keys, threshold=0.5):
@@ -543,31 +535,22 @@ class Agent(BaseModel):
                 best_match = key
         
         return best_match, best_score
-
     def _validate_response(self, response: Union[str, Dict]) -> Union[str, Dict]:
-        """Validate the response against the expected_output format using semantic similarity or fallback methods."""
+        """
+        Validate and structure the response based on the expected_output using semantic matching.
+        """
         if isinstance(self.expected_output, dict):
             if not isinstance(response, dict):
                 return {"response": response}
-
             validated_response = {}
             normalized_expected_keys = {self.normalize_key(k): k for k in self.expected_output.keys()}
-
             for expected_key_norm, expected_key_orig in normalized_expected_keys.items():
-                # Find all response keys that match the expected key (case-insensitive and normalized)
-                matching_response_keys = [
-                    k for k in response.keys()
-                    if self.normalize_key(k) == expected_key_norm
-                ]
-
-                # If no exact match, use semantic matching to find similar keys
+                matching_response_keys = [k for k in response.keys() if self.normalize_key(k) == expected_key_norm]
                 if not matching_response_keys:
                     for response_key in response.keys():
                         best_match, best_score = self.match_key(expected_key_orig, [response_key])
-                        if best_match and best_score > 0.5:  # Use a threshold to determine a valid match
+                        if best_match and best_score > 0.5:
                             matching_response_keys.append(response_key)
-
-                # Merge values from all matching keys
                 merged_values = []
                 for matching_key in matching_response_keys:
                     value = response[matching_key]
@@ -575,50 +558,41 @@ class Agent(BaseModel):
                         merged_values.extend(value)
                     else:
                         merged_values.append(value)
-
-                # Assign the merged values to the expected key
-                if merged_values:
-                    validated_response[expected_key_orig] = merged_values
-                else:
-                    validated_response[expected_key_orig] = "NA"  # Default value for missing keys
-
-                # Recursively validate nested dictionaries
+                validated_response[expected_key_orig] = merged_values if merged_values else "NA"
                 expected_value = self.expected_output[expected_key_orig]
                 if isinstance(expected_value, dict) and isinstance(validated_response[expected_key_orig], dict):
                     validated_response[expected_key_orig] = self._validate_response(validated_response[expected_key_orig])
-
             return validated_response
         elif isinstance(self.expected_output, str):
             if not isinstance(response, str):
                 return str(response)
         return response
-    
+
     def cli_app(
         self,
         message: Optional[str] = None,
         exit_on: Optional[List[str]] = None,
         **kwargs,
     ):
-        """Run the agent in a CLI app."""
+        """Run the agent as a command-line application."""
         from rich.prompt import Prompt
 
-        # Print initial message if provided
         if message:
             self.print_response(message=message, **kwargs)
 
         _exit_on = exit_on or ["exit", "quit", "bye"]
         while True:
             try:
-                message = Prompt.ask(f"[bold] {self.emoji} {self.user_name} [/bold]")
-                if message in _exit_on:
+                user_input = Prompt.ask(f"[bold] {self.emoji} {self.user_name} [/bold]")
+                if user_input in _exit_on:
                     break
-                self.print_response(message=message, **kwargs)
+                self.print_response(message=user_input, **kwargs)
             except KeyboardInterrupt:
                 print("\n\nSession ended. Goodbye!")
                 break
 
     def _generate_api(self):
-        """Generate an API for the agent if api=True."""
+        """Generate an API for the agent if API mode is enabled."""
         from .api.api_generator import APIGenerator
         self.api_generator = APIGenerator(self)
         print(f"API generated for agent '{self.name}'. Use `.run_api()` to start the API server.")
@@ -627,76 +601,7 @@ class Agent(BaseModel):
         """Run the API server for the agent."""
         if not hasattr(self, 'api_generator'):
             raise ValueError("API is not enabled for this agent. Set `api=True` when initializing the agent.")
-    
-        # Get API configuration
         host = self.api_config.get("host", "0.0.0.0") if self.api_config else "0.0.0.0"
         port = self.api_config.get("port", 8000) if self.api_config else 8000
-
-        # Run the API server
         self.api_generator.run(host=host, port=port)
 
-    def _flatten_data(self, data: Union[Dict, List], parent_key: str = "", separator: str = "_") -> List[Dict]:
-        """
-        Recursively flatten a nested dictionary or list into a list of key-value pairs.
-
-        Args:
-            data (Union[Dict, List]): The nested data structure.
-            parent_key (str): The parent key (used for recursion).
-            separator (str): The separator used for nested keys.
-
-        Returns:
-            List[Dict]: A list of flattened key-value pairs.
-        """
-        items = []
-        if isinstance(data, dict):
-            for key, value in data.items():
-                new_key = f"{parent_key}{separator}{key}" if parent_key else key
-                if isinstance(value, (dict, list)):
-                    items.extend(self._flatten_data(value, new_key, separator))
-                else:
-                    items.append({new_key: value})
-                    # Include the value as a key for searching
-                    if isinstance(value, str):
-                        items.append({value: new_key})
-        elif isinstance(data, list):
-            for index, item in enumerate(data):
-                new_key = f"{parent_key}{separator}{index}" if parent_key else str(index)
-                if isinstance(item, (dict, list)):
-                    items.extend(self._flatten_data(item, new_key, separator))
-                else:
-                    items.append({new_key: item})
-                    # Include the value as a key for searching
-                    if isinstance(item, str):
-                        items.append({item: new_key})
-        return items
-    
-    def _find_all_relevant_keys(self, query: str, flattened_data: List[Dict], threshold: float = 0.5) -> List[str]:
-        """
-        Find all relevant keys in the flattened data based on semantic similarity to the query.
-
-        Args:
-            query (str): The user's query.
-            flattened_data (List[Dict]): The flattened key-value pairs.
-            threshold (float): The similarity threshold for considering a match.
-
-        Returns:
-            List[str]: A list of relevant values.
-        """
-        if not flattened_data:
-            return []
-
-        # Extract keys from the flattened data
-        keys = [list(item.keys())[0] for item in flattened_data]
-
-        # Compute embeddings for the query and keys
-        query_embedding = self.semantic_model.encode(query, convert_to_tensor=True)
-        key_embeddings = self.semantic_model.encode(keys, convert_to_tensor=True)
-
-        # Compute cosine similarity between the query and keys
-        similarities = util.pytorch_cos_sim(query_embedding, key_embeddings)[0]
-
-        # Find all keys with a similarity score above the threshold
-        relevant_indices = [i for i, score in enumerate(similarities) if score > threshold]
-        relevant_values = [flattened_data[i][keys[i]] for i in relevant_indices]
-
-        return relevant_values

semantio/knowledge_base/document_loader.py

@@ -1,61 +1,191 @@
-from typing import List, Dict, Any
+import os
+import json
+import csv
+import re
 from pathlib import Path
+from typing import List, Dict, Any
+from io import BytesIO
 
-class DocumentLoader:
-    """
-    A class to load documents from various sources (e.g., files, URLs) into the knowledge base.
-    """
+import requests
+from bs4 import BeautifulSoup
 
-    def __init__(self):
-        """
-        Initialize the DocumentLoader.
-        """
-        pass
+# Optional: Import pandas for XLSX support and PyPDF2 for PDF support
+try:
+    import pandas as pd
+except ImportError:
+    pd = None
 
-    def load_from_file(self, file_path: str) -> List[Dict[str, Any]]:
-        """
-        Load documents from a file.
+try:
+    from PyPDF2 import PdfReader
+except ImportError:
+    PdfReader = None
 
-        Args:
-            file_path (str): The path to the file.
 
-        Returns:
-            List[Dict[str, Any]]: A list of documents, where each document is a dictionary.
-        """
-        file_path = Path(file_path)
-        if not file_path.exists():
-            raise FileNotFoundError(f"File not found: {file_path}")
+def flatten_json(data: Any, parent_key: str = "", separator: str = "_") -> List[Dict[str, Any]]:
+    """
+    Recursively flatten a JSON structure.
+    For each key-value pair, add an entry mapping key->value.
+    Additionally, if the value is a string, add an entry mapping the value to its flattened key.
+    """
+    items = []
+    if isinstance(data, dict):
+        for key, value in data.items():
+            new_key = f"{parent_key}{separator}{key}" if parent_key else key
+            if isinstance(value, (dict, list)):
+                items.extend(flatten_json(value, new_key, separator))
+            else:
+                items.append({new_key: value})
+                if isinstance(value, str):
+                    items.append({value: new_key})
+    elif isinstance(data, list):
+        for index, item in enumerate(data):
+            new_key = f"{parent_key}{separator}{index}" if parent_key else str(index)
+            if isinstance(item, (dict, list)):
+                items.extend(flatten_json(item, new_key, separator))
+            else:
+                items.append({new_key: item})
+                if isinstance(item, str):
+                    items.append({item: new_key})
+    return items
 
-        # Example: Load a JSON file
-        if file_path.suffix == ".json":
-            import json
-            with open(file_path, "r") as f:
-                return json.load(f)
-        # Example: Load a text file
-        elif file_path.suffix == ".txt":
-            with open(file_path, "r") as f:
-                return [{"text": f.read()}]
-        else:
-            raise ValueError(f"Unsupported file type: {file_path.suffix}")
 
-    def load_from_url(self, url: str) -> List[Dict[str, Any]]:
+class DocumentLoader:
+    """
+    A dynamic document loader that supports multiple source types:
+    
+    - Local files: CSV, TXT, JSON, XLSX, PDF
+    - URL sources: HTML websites (text extraction), JSON APIs, PDF URLs
+    - YouTube links: Extracts transcripts using youtube_transcript_api
+    
+    For JSON sources, if flatten is True (default), the returned document is a dictionary with two keys:
+       "original": the raw JSON data,
+       "flattened": a list of flattened key/value pairs (including reverse mappings).
+    """
+    def load(self, source: str, flatten: bool = True) -> List[Dict[str, Any]]:
         """
-        Load documents from a URL.
+        Load documents from the given source.
+        If source starts with "http", treat it as a URL; otherwise, as a local file.
+        """
+        if source.startswith("http"):
+            return self.load_from_url(source, flatten=flatten)
+        else:
+            return self.load_from_file(source, flatten=flatten)
 
-        Args:
-            url (str): The URL to load documents from.
+    def load_from_file(self, file_path: str, flatten: bool = True) -> List[Dict[str, Any]]:
+        path = Path(file_path)
+        if not path.exists():
+            raise FileNotFoundError(f"File not found: {file_path}")
+        ext = path.suffix.lower()
+        if ext == ".json":
+            with open(path, "r", encoding="utf-8") as f:
+                data = json.load(f)
+            if flatten:
+                return [{"original": data, "flattened": flatten_json(data)}]
+            else:
+                return data if isinstance(data, list) else [data]
+        elif ext == ".txt":
+            with open(path, "r", encoding="utf-8") as f:
+                content = f.read()
+            return [{"text": content}]
+        elif ext == ".csv":
+            with open(path, "r", encoding="utf-8") as f:
+                reader = csv.DictReader(f)
+                return [row for row in reader]
+        elif ext == ".xlsx":
+            if pd is None:
+                raise ImportError("pandas is required to load XLSX files")
+            df = pd.read_excel(path)
+            return df.to_dict(orient="records")
+        elif ext == ".pdf":
+            if PdfReader is None:
+                raise ImportError("PyPDF2 is required to load PDF files")
+            reader = PdfReader(str(path))
+            content = ""
+            for page in reader.pages:
+                content += page.extract_text() or ""
+            return [{"text": content}]
+        else:
+            raise ValueError(f"Unsupported file type: {ext}")
 
-        Returns:
-            List[Dict[str, Any]]: A list of documents, where each document is a dictionary.
-        """
-        import requests
+    def load_from_url(self, url: str, flatten: bool = True) -> List[Dict[str, Any]]:
+        if "youtube.com" in url or "youtu.be" in url:
+            return self._load_youtube(url)
         response = requests.get(url)
         if response.status_code != 200:
             raise ValueError(f"Failed to fetch data from URL: {url}")
-
-        # Example: Load JSON data from a URL
-        if "application/json" in response.headers.get("Content-Type", ""):
-            return response.json()
-        # Example: Load text data from a URL
+        content_type = response.headers.get("Content-Type", "").lower()
+        if "application/json" in content_type:
+            data = response.json()
+            if flatten:
+                return [{"original": data, "flattened": flatten_json(data)}]
+            else:
+                return data if isinstance(data, list) else [data]
+        elif "text/html" in content_type:
+            # First, try with requests + BeautifulSoup.
+            soup = BeautifulSoup(response.text, "html.parser")
+            text = soup.get_text(separator="\n").strip()
+            # If the text seems too short (less than 50 words), assume content is loaded via JavaScript.
+            if len(text.split()) < 50:
+                try:
+                    text = self._fetch_with_headless_browser(url)
+                except Exception as e:
+                    # If headless browser fails, log and fallback to the short text.
+                    print(f"Headless fetch failed: {e}")
+            return [{"text": text}]
+        elif "application/pdf" in content_type:
+            if PdfReader is None:
+                raise ImportError("PyPDF2 is required to load PDF files")
+            pdf_file = BytesIO(response.content)
+            reader = PdfReader(pdf_file)
+            text = ""
+            for page in reader.pages:
+                text += page.extract_text() or ""
+            return [{"text": text}]
         else:
-            return [{"text": response.text}]
+            return [{"text": response.text}]
+
+    def _fetch_with_headless_browser(self, url: str) -> str:
+        """
+        Use a headless browser (Playwright) to fetch fully rendered content.
+        """
+        try:
+            from playwright.sync_api import sync_playwright
+        except ImportError:
+            raise ImportError("playwright is required for JS-rendered pages. Install it with 'pip install playwright' and run 'playwright install'.")
+        with sync_playwright() as p:
+            browser = p.chromium.launch(headless=True)
+            page = browser.new_page()
+            page.goto(url, wait_until="networkidle")
+            html = page.content()
+            browser.close()
+            soup = BeautifulSoup(html, "html.parser")
+            text = soup.get_text(separator="\n").strip()
+            return text
+
+    def _load_youtube(self, url: str) -> List[Dict[str, Any]]:
+        try:
+            from youtube_transcript_api import YouTubeTranscriptApi
+        except ImportError:
+            raise ImportError("youtube_transcript_api is required to load YouTube transcripts")
+        
+        video_id = None
+        patterns = [r"v=([^&]+)", r"youtu\.be/([^?&]+)"]
+        for pattern in patterns:
+            match = re.search(pattern, url)
+            if match:
+                video_id = match.group(1)
+                break
+        if not video_id:
+            raise ValueError("Could not extract video ID from URL")
+        
+        # Define a prioritized list of language codes to try
+        preferred_languages = ["en", "hi", "es", "fr", "de", "ru"]
+        
+        try:
+            transcript = YouTubeTranscriptApi.get_transcript(video_id, languages=preferred_languages)
+            text = " ".join(segment["text"] for segment in transcript)
+            return [{"text": text}]
+        except Exception as e:
+            # Return a fallback document indicating transcript retrieval failed
+            return [{"text": f"Transcript not available for video {url}: {str(e)}"}]
+

{semantio-0.0.7.dist-info → semantio-0.0.8.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: semantio
-Version: 0.0.7
+Version: 0.0.8
 Summary: A powerful SDK for building AI agents
 Home-page: https://github.com/Syenah/semantio
 Author: Rakesh
@@ -36,6 +36,9 @@ Requires-Dist: yfinance
 Requires-Dist: beautifulsoup4
 Requires-Dist: webdriver-manager
 Requires-Dist: validators
+Requires-Dist: PyPDF2
+Requires-Dist: youtube-transcript-api
+Requires-Dist: pandas
 
 # Semantio: The Mother of Your AI Agents
 

{semantio-0.0.7.dist-info → semantio-0.0.8.dist-info}/RECORD

@@ -1,5 +1,5 @@
 semantio/__init__.py,sha256=RIeSI07dGyWBK-STKIk4IeB4bkn_3-QEKQklzSvR7hQ,82
-semantio/agent.py,sha256=uPFz1WP2eb-z-tryQOX8necS8_tv4Il6qxNmZux9hNk,31709
+semantio/agent.py,sha256=ND-EBsY4vRgmmmooVjz3iRR-8VI8Z7A14-xg_5c18Ho,28060
 semantio/memory.py,sha256=en9n3UySnj4rA0x3uR1sEdEzA7EkboQNbEHQ5KuEehw,2115
 semantio/models.py,sha256=7hmP-F_aSU8WvsG3NGeC_hep-rUbiSbjUFMDVbpKxQE,289
 semantio/rag.py,sha256=ROy3Pa1NURcDs6qQZ8IMoa5Xlzt6I-msEq0C1p8UgB0,472
@@ -9,7 +9,7 @@ semantio/api/fastapi_app.py,sha256=DyTgKJKikMe2G6wWmyzo1rBLXQFi8UWWUMY3UGH4f24,2
 semantio/cli/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 semantio/cli/main.py,sha256=jUvSfehbHWALwracEgBopMIVMraSV9QmDUFfgGcxnP0,1091
 semantio/knowledge_base/__init__.py,sha256=mvp0GFiGSjcxlkaDulAwKOCL9s6gsKTqhPKXF9N3n1g,172
-semantio/knowledge_base/document_loader.py,sha256=nix0yZJ-JJoDbhLkpg5bKDMvNrwykmknI7MRIn0N81k,1910
+semantio/knowledge_base/document_loader.py,sha256=g7a09hxAZRdwXS8JG__0ZXf1Yh4sa-xW2NqLNEMWwXM,7822
 semantio/knowledge_base/retriever.py,sha256=XpdzKS1UCncJImVMtG67VXMC7lp2eRzKnShjvktsFMM,1271
 semantio/knowledge_base/vector_store.py,sha256=4Zv9kfqDD3cfn_4R8ZoLKdAQCZRYo_IENP_KkLB_RPc,987
 semantio/llm/__init__.py,sha256=-4uKcqo9fBrEbvfxGE01XVHL9qEG2vKXfy5hlnUsRbw,779
@@ -37,9 +37,9 @@ semantio/utils/date_utils.py,sha256=x3oqRGv6ee_KCJ0LvCqqZh_FSgS6YGOHBwZQS4TJetY,
 semantio/utils/file_utils.py,sha256=b_cMuJINEGk9ikNuNHSn9lsmICWwvtnCDZ03ndH_S2I,1779
 semantio/utils/logger.py,sha256=TmGbP8BRjLMWjXi2GWzZ0RIXt70x9qX3FuIqghCNlwM,510
 semantio/utils/validation_utils.py,sha256=iwoxEb4Q5ILqV6tbesMjPWPCCoL3AmPLejGUy6q8YvQ,1284
-semantio-0.0.7.dist-info/LICENSE,sha256=mziLlfb9hZ8HKxm9V6BiHpmgJvmcDvswu1QBlDB-6vU,1074
-semantio-0.0.7.dist-info/METADATA,sha256=QQRzinLKReosRRthYf1bei5FDAaOPHaG4bG5gdJnMFc,6889
-semantio-0.0.7.dist-info/WHEEL,sha256=ewwEueio1C2XeHTvT17n8dZUJgOvyCWCt0WVNLClP9o,92
-semantio-0.0.7.dist-info/entry_points.txt,sha256=zbPgevSLwcLpdRHqI_atE8EOt8lK2vRF1AoDflDTo18,53
-semantio-0.0.7.dist-info/top_level.txt,sha256=Yte_6mb-bh-I_lQwMjk1GijZkxPoX4Zmp3kBftC1ZlA,9
-semantio-0.0.7.dist-info/RECORD,,
+semantio-0.0.8.dist-info/LICENSE,sha256=mziLlfb9hZ8HKxm9V6BiHpmgJvmcDvswu1QBlDB-6vU,1074
+semantio-0.0.8.dist-info/METADATA,sha256=et3Zs5Q_F-izo_We1M1gbykDVRNud7Qk_7uTDkNBjkw,6971
+semantio-0.0.8.dist-info/WHEEL,sha256=ewwEueio1C2XeHTvT17n8dZUJgOvyCWCt0WVNLClP9o,92
+semantio-0.0.8.dist-info/entry_points.txt,sha256=zbPgevSLwcLpdRHqI_atE8EOt8lK2vRF1AoDflDTo18,53
+semantio-0.0.8.dist-info/top_level.txt,sha256=Yte_6mb-bh-I_lQwMjk1GijZkxPoX4Zmp3kBftC1ZlA,9
+semantio-0.0.8.dist-info/RECORD,,