hana-ai 1.0.25063000__tar.gz → 1.0.25070200__tar.gz

{hana_ai-1.0.25063000/src/hana_ai.egg-info → hana_ai-1.0.25070200}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hana_ai
-Version: 1.0.25063000
+Version: 1.0.25070200
 Summary: Generative AI Toolkit for SAP HANA Cloud
 Author: SAP
 License: Apache-2.0
@@ -13,7 +13,7 @@ Requires-Python: >=3.0
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: langchain
-Requires-Dist: numpy<2.0
+Requires-Dist: numpy
 Requires-Dist: pandas
 Requires-Dist: hana-ml>=2.25.25062301
 Requires-Dist: langchain-community
@@ -23,7 +23,9 @@ Requires-Dist: langchain-text-splitters
 Requires-Dist: pydantic
 Requires-Dist: pydantic-core
 Requires-Dist: cryptography
-Requires-Dist: ipython<9.0
+Requires-Dist: ipython
+Requires-Dist: sentence_transformers
+Requires-Dist: deprecated
 Dynamic: license-file
 
 [![REUSE status](https://api.reuse.software/badge/github.com/SAP/generative-ai-toolkit-for-sap-hana-cloud)](https://api.reuse.software/info/github.com/SAP/generative-ai-toolkit-for-sap-hana-cloud)

{hana_ai-1.0.25063000 → hana_ai-1.0.25070200}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "hana_ai"
-version = "1.0.25063000"
+version = "1.0.25070200"
 authors = [
     {name = "SAP"}
 ]

{hana_ai-1.0.25063000 → hana_ai-1.0.25070200}/requirements.txt

@@ -1,5 +1,5 @@
 langchain
-numpy<2.0
+numpy
 pandas
 hana-ml>=2.25.25062301
 langchain-community
@@ -9,4 +9,6 @@ langchain-text-splitters
 pydantic
 pydantic-core
 cryptography
-ipython<9.0
+ipython
+sentence_transformers
+deprecated

{hana_ai-1.0.25063000 → hana_ai-1.0.25070200}/src/hana_ai/__init__.py

@@ -2,4 +2,4 @@
 hana_ai is a Python package for AI/ML related utilities.
 """
 
-__version__ = "1.0.25063000"
+__version__ = "1.0.25070200"

{hana_ai-1.0.25063000 → hana_ai-1.0.25070200}/src/hana_ai/agents/hanaml_agent_with_memory.py

@@ -11,6 +11,7 @@ The following class is available:
 import json
 import inspect
 import logging
+from deprecated import deprecated
 import pandas as pd
 from pydantic import ValidationError
 from langchain.agents import initialize_agent, AgentType, Tool
@@ -124,6 +125,8 @@ def _get_pandas_meta(df):
         columns = df.columns.tolist()
         return json.dumps({"columns": columns})
     return ''
+
+@deprecated(version="1.0.25070200", reason="Use HANAMLRagAgent instead")
 class HANAMLAgentWithMemory(object):
     """
     A chatbot that can remember the chat history and use it to generate responses.

hana_ai-1.0.25070200/src/hana_ai/agents/hanaml_rag_agent.py

@@ -0,0 +1,367 @@
+"""
+A chatbot that can remember the short term and long term chat history and use it to generate responses.
+
+The following class is available:
+    
+        * :class `HANAMLRAGAgent`
+"""
+
+import re
+from typing import List, Any
+from datetime import datetime
+import logging
+import pandas as pd
+from langchain_text_splitters import RecursiveCharacterTextSplitter
+
+from langchain.agents import Tool, AgentExecutor, create_openai_functions_agent
+from langchain.memory import ConversationBufferWindowMemory
+from langchain.prompts import ChatPromptTemplate, MessagesPlaceholder, HumanMessagePromptTemplate
+from langchain.schema import SystemMessage, HumanMessage, AIMessage
+from langchain_community.chat_message_histories import SQLChatMessageHistory
+from langchain_community.vectorstores import FAISS
+
+from sentence_transformers import CrossEncoder
+from hana_ai.vectorstore.embedding_service import GenAIHubEmbeddings
+
+# Configure logging
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+class HANAMLRAGAgent:
+    """
+    A chatbot that integrates short-term and long-term memory systems using RAG (Retrieval-Augmented Generation).
+    """
+    def __init__(self,
+                 tools: List[Tool],
+                 llm: Any,
+                 vectorstore_path: str = "chat_history_vectorstore",
+                 memory_window: int = 10,
+                 long_term_db: str = "sqlite:///chat_history.db",
+                 long_term_memory_limit: int = 1000,
+                 skip_large_data_threshold: int = 1000,
+                 chunk_size: int = 500,
+                 chunk_overlap: int = 50,
+                 forget_percentage: float = 0.1,
+                 max_iterations: int = 10,
+                 cross_encoder: CrossEncoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2'),
+                 rerank_k: int = 10,
+                 score_threshold: float = 0.5,
+                 verbose: bool = False):
+        """
+        Initialize the chatbot with integrated tools and memory systems.
+
+        Parameters
+        ----------
+        tools : List[Tool]
+            List of LangChain tools to be used by the agent.
+        llm : Any
+            Language model instance to be used for generating responses.
+        vectorstore_path : str
+            Path to store the vectorstore for long-term memory.
+        memory_window : int
+            Number of recent conversations to keep in short-term memory.
+
+            Defaults to 10.
+        long_term_db : str
+            Connection string for long-term memory storage.
+
+            Defaults to "sqlite:///chat_history.db".
+        long_term_memory_limit : int
+            Maximum number of long-term memory entries to retain.
+
+            Defaults to 1000.
+        skip_large_data_threshold : int
+            Skip storing texts longer than this threshold.
+
+            Defaults to 1000.
+        chunk_size : int
+            Text chunk size for embeddings.
+
+            Defaults to 500.
+        chunk_overlap : int
+            Text chunk overlap for embeddings.
+
+            Defaults to 50.
+        forget_percentage : float
+            Percentage of oldest memories to forget when long-term memory limit is reached.
+
+            Defaults to 0.1 (10%).
+        max_iterations : int
+            Maximum number of iterations for agent execution.
+
+            Defaults to 10.
+        cross_encoder : CrossEncoder
+            Cross-encoder model for reranking retrieved documents.
+
+            Defaults to a pre-trained cross-encoder model 'cross-encoder/ms-marco-MiniLM-L-6-v2'.            
+        rerank_k : int
+            Number of documents to retrieve for reranking.
+
+            Defaults to 10.
+        score_threshold : float
+            Similarity score threshold for retrieval.
+
+            Defaults to 0.5.
+        verbose : bool
+            Whether to enable verbose logging.
+
+            Defaults to False.
+        """
+        self.llm = llm
+        self.tools = tools
+        self.vectorstore_path = vectorstore_path
+        self.long_term_db = long_term_db
+        self.long_term_memory_limit = long_term_memory_limit
+        self.memory_window = memory_window
+        self.skip_large_data_threshold = skip_large_data_threshold
+        self.chunk_size = chunk_size
+        self.chunk_overlap = chunk_overlap
+        self.forget_percentage = forget_percentage
+        self.max_iterations = max_iterations
+        self.rerank_k = rerank_k
+        self.score_threshold = score_threshold
+        self.verbose = verbose
+        self.cross_encoder = cross_encoder
+
+        # Initialize memory systems
+        self._initialize_memory()
+
+        # Initialize agent with tool integration
+        self._initialize_agent()
+
+    def _initialize_memory(self):
+        """Initialize short-term and long-term memory systems"""
+        # Short-term memory (recent conversations)
+        self.short_term_memory = ConversationBufferWindowMemory(
+            memory_key="chat_history",
+            k=self.memory_window,
+            return_messages=True
+        )
+        # Long-term memory storage
+        self.long_term_store = SQLChatMessageHistory(
+            connection_string=self.long_term_db,
+            session_id="global_session"
+        )
+        # Initialize RAG vectorstore
+        self._initialize_vectorstore()
+
+    def _initialize_vectorstore(self):
+        """Initialize or load FAISS vectorstore for long-term memory"""
+        # Create embeddings service
+        self.embeddings = GenAIHubEmbeddings()
+
+        # Try loading existing vectorstore
+        try:
+            self.vectorstore = FAISS.load_local(
+                self.vectorstore_path,
+                self.embeddings,
+                allow_dangerous_deserialization=True
+            )
+            logger.info("Loaded existing vectorstore from %s", self.vectorstore_path)
+        except:
+            # Initialize new vectorstore if loading fails
+            self.vectorstore = FAISS.from_texts(
+                texts=["Initialization text"],
+                embedding=self.embeddings
+            )
+            self.vectorstore.save_local(self.vectorstore_path)
+            logger.info("Created new vectorstore at %s", self.vectorstore_path)
+
+    def _should_store(self, text: str) -> bool:
+        """Determine if text should be stored in memory"""
+        # Skip DataFrame conversions
+        if isinstance(text, pd.DataFrame):
+            return False
+
+        # Skip large texts
+        if isinstance(text, str) and len(text) > self.skip_large_data_threshold:
+            return False
+
+        # Skip Markdown tables
+        if isinstance(text, str) and re.search(r'\|\s*.*\s*\|.*\n(\|\s*:?[-]+\s*:?\|\s*)+\n', text):
+            return False
+
+        return True
+
+    def _update_long_term_memory(self, user_input: str, response: Any):
+        """Update long-term memory with new conversation"""
+        if not self._should_store(response):
+            logger.debug("Skipping memory storage for large or special content")
+            return
+
+        response_str = str(response)
+        current_time = datetime.now().isoformat()
+
+        # Add messages with metadata
+        self.long_term_store.add_messages([
+            HumanMessage(content=user_input, metadata={"timestamp": current_time}),
+            AIMessage(content=response_str, metadata={"timestamp": current_time})
+        ])
+
+        # Create documents for vector store
+        splitter = RecursiveCharacterTextSplitter(
+            chunk_size=self.chunk_size,
+            chunk_overlap=self.chunk_overlap,
+        )
+
+        documents = splitter.create_documents(
+            texts=[f"User: {user_input}\nAssistant: {response_str}"],
+            metadatas=[{"timestamp": current_time}]
+        )
+
+        # Update vector store
+        self.vectorstore.add_documents(documents)
+        self.vectorstore.save_local(self.vectorstore_path)
+
+        # Clean up oldest memories if needed
+        self._forget_old_memories()
+
+    def _forget_old_memories(self):
+        """Remove oldest memories when storage limit is exceeded"""
+        if len(self.long_term_store.messages) > self.long_term_memory_limit:
+            # Calculate number of oldest messages to remove
+            num_to_remove = int(self.long_term_memory_limit * self.forget_percentage)
+
+            # Implement actual deletion logic here
+            # This is placeholder for actual database deletion
+            logger.info("Should remove %d oldest memories", num_to_remove)
+            # In practice: Delete oldest records from database
+
+    def _retrieve_relevant_memories(self, query: str) -> List[str]:
+        """Retrieve relevant memories using RAG with reranking"""
+        # Retrieve candidate documents
+        candidate_docs = self.vectorstore.similarity_search_with_score(
+            query=query,
+            k=self.rerank_k,
+            score_threshold=self.score_threshold
+        )
+
+        if not candidate_docs:
+            return []
+
+        # Prioritize recent documents
+        candidate_docs.sort(
+            key=lambda x: x[0].metadata.get('timestamp', '1970-01-01'),
+            reverse=True
+        )
+
+        # Rerank with cross-encoder
+        doc_contents = [doc[0].page_content for doc in candidate_docs]
+        rerank_scores = self.cross_encoder.predict([(query, content) for content in doc_contents])
+
+        # Combine and select top documents
+        combined = list(zip(doc_contents, rerank_scores))
+        combined.sort(key=lambda x: x[1], reverse=True)
+
+        return [content for content, _ in combined[:3]]
+
+    def _build_context(self, user_input: str) -> ChatPromptTemplate:
+        """Build combined context from short-term and long-term memory"""
+        # Get short-term memory
+        short_term_history = self.short_term_memory.load_memory_variables({})["chat_history"]
+        short_term_str = "\n".join([
+            f"{msg.type.capitalize()}: {msg.content}"
+            for msg in short_term_history
+        ]) if short_term_history else "No recent conversation history."
+
+        # Get relevant long-term memories
+        long_term_context = self._retrieve_relevant_memories(user_input)
+        long_term_str = "\n".join(long_term_context) if long_term_context else "No relevant historical context found."
+
+        # Create system message with context
+        context_message = SystemMessage(
+            content=f"CONVERSATION CONTEXT:\n"
+                    f"Recent Chat History:\n{short_term_str}\n\n"
+                    f"Relevant Long-term Memories:\n{long_term_str}\n\n"
+                    "GUIDELINES:\n"
+                    "1. Prioritize tool usage when appropriate\n"
+                    "2. When tools return tabular data, format as Markdown tables\n"
+                    "3. For large outputs, summarize key insights\n"
+        )
+
+        return context_message
+
+    def _initialize_agent(self):
+        """Initialize agent with tool integration and context handling"""
+        # Create the prompt template
+        prompt = ChatPromptTemplate.from_messages([
+            SystemMessage(content="You are a helpful assistant with access to tools. Always use tools when appropriate."),
+            MessagesPlaceholder(variable_name="chat_history"),
+            HumanMessagePromptTemplate.from_template("{input}"),
+            MessagesPlaceholder(variable_name="agent_scratchpad")
+        ])
+
+        # Create the agent with tool bindings
+        self.agent = create_openai_functions_agent(self.llm, self.tools, prompt)
+
+        # Create executor with memory integration
+        self.executor = AgentExecutor(
+            agent=self.agent,
+            tools=self.tools,
+            max_iterations=self.max_iterations,
+            verbose=self.verbose,
+            memory=self.short_term_memory,
+            handle_parsing_errors=True
+        )
+
+    def _format_dataframe(self, df: pd.DataFrame) -> str:
+        """Convert DataFrame to Markdown table"""
+        try:
+            return df.to_markdown(index=False)
+        except Exception as e:
+            logger.error("DataFrame conversion failed: %s", str(e))
+            return "Data output could not be formatted"
+
+    def _build_long_term_context(self, user_input: str) -> str:
+        long_term_context = self._retrieve_relevant_memories(user_input)
+        return "\n".join(long_term_context) if long_term_context else ""
+
+    def clear_long_term_memory(self):
+        """
+        Safely clear long-term memory by clearing the vectorstore and database.
+        Avoids index errors in SQLChatMessageHistory implementation.
+        """
+        try:
+            # 安全清空SQL存储（避免索引越界）
+            self.long_term_store.clear()
+            logger.debug("SQL long-term store cleared")
+        except IndexError:
+            # 捕获索引越界异常（空存储时可能发生）
+            logger.warning("IndexError during clear (likely empty store)")
+        except Exception as e:
+            logger.error("Unexpected error clearing SQL store: %s", str(e))
+
+        # 重建空向量库（无需前置检查）
+        self.vectorstore = FAISS.from_texts([""], embedding=self.embeddings)  # 空文本占位
+        self.vectorstore.save_local(self.vectorstore_path)
+        logger.info("Long-term memory cleared and vectorstore reset.")
+
+    def clear_short_term_memory(self):
+        """
+        Clear short-term memory by resetting the conversation history.
+        This is a placeholder for actual implementation.
+        """
+        self.short_term_memory.clear()
+        logger.info("Short-term memory cleared.")
+
+    def chat(self, user_input: str) -> str:
+        """
+        Main chat method to handle user input and return response.
+
+        Parameters
+        ----------
+        user_input : str
+            The input question or statement from the user.
+        """
+        if user_input.startswith("!clear_long_term_memory"):
+            self.clear_long_term_memory()
+            return "Long-term memory has been cleared."
+        elif user_input.startswith("!clear_short_term_memory"):
+            self.clear_short_term_memory()
+            return "Short-term memory has been cleared."
+        context_str = self._build_long_term_context(user_input)  # Returns string
+        combined_input = f"Context:\n{context_str}\n\nQuestion: {user_input}"
+
+        agent_input = {"input": combined_input}  # Single key
+        response = self.executor.invoke(agent_input)
+        return response['output']

{hana_ai-1.0.25063000 → hana_ai-1.0.25070200}/src/hana_ai/vectorstore/embedding_service.py

@@ -20,7 +20,7 @@ import sys
 try:
     from gen_ai_hub.proxy.langchain import init_embedding_model as gen_ai_hub_embedding_model
 except ImportError:
-    subprocess.check_call([sys.executable, "-m", "pip", "install", "generative-ai-hub-sdk[all]"])
+    subprocess.check_call([sys.executable, "-m", "pip", "install", "sap-ai-sdk-gen[all]"])
     from gen_ai_hub.proxy.langchain import init_embedding_model as gen_ai_hub_embedding_model
 
 import pandas as pd

{hana_ai-1.0.25063000 → hana_ai-1.0.25070200/src/hana_ai.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hana_ai
-Version: 1.0.25063000
+Version: 1.0.25070200
 Summary: Generative AI Toolkit for SAP HANA Cloud
 Author: SAP
 License: Apache-2.0
@@ -13,7 +13,7 @@ Requires-Python: >=3.0
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: langchain
-Requires-Dist: numpy<2.0
+Requires-Dist: numpy
 Requires-Dist: pandas
 Requires-Dist: hana-ml>=2.25.25062301
 Requires-Dist: langchain-community
@@ -23,7 +23,9 @@ Requires-Dist: langchain-text-splitters
 Requires-Dist: pydantic
 Requires-Dist: pydantic-core
 Requires-Dist: cryptography
-Requires-Dist: ipython<9.0
+Requires-Dist: ipython
+Requires-Dist: sentence_transformers
+Requires-Dist: deprecated
 Dynamic: license-file
 
 [![REUSE status](https://api.reuse.software/badge/github.com/SAP/generative-ai-toolkit-for-sap-hana-cloud)](https://api.reuse.software/info/github.com/SAP/generative-ai-toolkit-for-sap-hana-cloud)

{hana_ai-1.0.25063000 → hana_ai-1.0.25070200}/src/hana_ai.egg-info/SOURCES.txt

@@ -32,6 +32,7 @@ src/hana_ai/agents/hana_dataframe_agent.py
 src/hana_ai/agents/hana_dataframe_prompt.py
 src/hana_ai/agents/hana_sql_agent.py
 src/hana_ai/agents/hanaml_agent_with_memory.py
+src/hana_ai/agents/hanaml_rag_agent.py
 src/hana_ai/tools/__init__.py
 src/hana_ai/tools/agent_as_a_tool.py
 src/hana_ai/tools/code_template_tools.py

{hana_ai-1.0.25063000 → hana_ai-1.0.25070200}/src/hana_ai.egg-info/requires.txt

@@ -1,5 +1,5 @@
 langchain
-numpy<2.0
+numpy
 pandas
 hana-ml>=2.25.25062301
 langchain-community
@@ -9,4 +9,6 @@ langchain-text-splitters
 pydantic
 pydantic-core
 cryptography
-ipython<9.0
+ipython
+sentence_transformers
+deprecated