alita-sdk 0.3.263__py3-none-any.whl → 0.3.499__py3-none-any.whl

alita_sdk/runtime/tools/sandbox.py

@@ -0,0 +1,375 @@
+import asyncio
+import logging
+import subprocess
+import os
+from typing import Any, Type, Optional, Dict, List, Literal, Union
+from copy import deepcopy
+from pathlib import Path
+
+from langchain_core.tools import BaseTool, BaseToolkit
+from langchain_core.messages import ToolCall
+from pydantic import BaseModel, create_model, ConfigDict, Field
+from pydantic.fields import FieldInfo
+
+logger = logging.getLogger(__name__)
+
+name = "pyodide"
+
+
+def get_tools(tools_list: list, alita_client=None, llm=None, memory_store=None):
+    """
+    Get sandbox tools for the provided tool configurations.
+
+    Args:
+        tools_list: List of tool configurations
+        alita_client: Alita client instance for sandbox tools
+        llm: LLM client instance (unused for sandbox)
+        memory_store: Optional memory store instance (unused for sandbox)
+
+    Returns:
+        List of sandbox tools
+    """
+    all_tools = []
+
+    for tool in tools_list:
+        if tool.get('type') == 'sandbox' or tool.get('toolkit_name') == 'sandbox':
+            try:
+                toolkit_instance = SandboxToolkit.get_toolkit(
+                    stateful=tool['settings'].get('stateful', False),
+                    allow_net=tool['settings'].get('allow_net', True),
+                    alita_client=alita_client,
+                    toolkit_name=tool.get('toolkit_name', '')
+                )
+                all_tools.extend(toolkit_instance.get_tools())
+            except Exception as e:
+                logger.error(f"Error in sandbox toolkit get_tools: {e}")
+                logger.error(f"Tool config: {tool}")
+                raise
+
+    return all_tools
+
+
+def _is_deno_available() -> bool:
+    """Check if Deno is available in the PATH"""
+    try:
+        result = subprocess.run(
+            ["deno", "--version"],
+            capture_output=True,
+            text=True,
+            timeout=10
+        )
+        return result.returncode == 0
+    except (subprocess.TimeoutExpired, subprocess.CalledProcessError, FileNotFoundError):
+        return False
+
+
+def _setup_pyodide_cache_env() -> None:
+    """Setup Pyodide caching environment variables for performance optimization [NO-OP]"""
+    try:
+        for key in ["SANDBOX_BASE", "DENO_DIR"]:
+            logger.info("Sandbox env: %s -> %s", key, os.environ.get(key, "n/a"))
+    except Exception as e:
+        logger.warning(f"Could not setup Pyodide cache environment: {e}")
+
+
+# Create input schema for the sandbox tool
+sandbox_tool_input = create_model(
+    "SandboxToolInput",
+    code=(str, FieldInfo(description="Python code to execute in the sandbox environment"))
+)
+
+
+class PyodideSandboxTool(BaseTool):
+    """
+    A tool that provides secure Python code execution using Pyodide (Python compiled to WebAssembly).
+    This tool leverages langchain-sandbox to provide a safe environment for running untrusted Python code.
+    Optimized for performance with caching and stateless execution by default.
+    """
+
+    name: str = "pyodide_sandbox"
+    description: str = """Execute Python code in a secure sandbox environment using Pyodide.
+    This tool allows safe execution of Python code without access to the host system.
+    Use this tool when you need to:
+    - Execute Python code snippets
+    - Perform calculations or data analysis
+    - Test Python algorithms
+    - Run code that requires isolation from the host system
+
+    The sandbox supports most Python standard library modules and can install additional packages.
+    Note: File access and some system operations are restricted for security.
+    Optimized for performance with local caching (stateless by default for faster execution).
+    """
+    args_schema: Type[BaseModel] = sandbox_tool_input
+    stateful: bool = False  # Default to stateless for better performance
+    allow_net: bool = True
+    session_bytes: Optional[bytes] = None
+    session_metadata: Optional[Dict] = None
+    alita_client: Optional[Any] = None
+
+    def __init__(self, **kwargs: Any) -> None:
+        super().__init__(**kwargs)
+        self._sandbox = None
+        # Setup caching environment for optimal performance
+        _setup_pyodide_cache_env()
+        self._initialize_sandbox()
+
+    def _prepare_pyodide_input(self, code: str) -> str:
+        """Prepare input for PyodideSandboxTool by injecting state and alita_client into the code block."""
+        pyodide_predata = ""
+
+        # Add alita_client if available
+        if self.alita_client:
+            try:
+                # Get the directory of the current file and construct the path to sandbox_client.py
+                current_dir = Path(__file__).parent
+                sandbox_client_path = current_dir.parent / 'clients' / 'sandbox_client.py'
+
+                with open(sandbox_client_path, 'r') as f:
+                    sandbox_client_code = f.read()
+                pyodide_predata += f"{sandbox_client_code}\n"
+                pyodide_predata += (f"alita_client = SandboxClient(base_url='{self.alita_client.base_url}',"
+                                    f"project_id={self.alita_client.project_id},"
+                                    f"auth_token='{self.alita_client.auth_token}')\n")
+            except FileNotFoundError:
+                logger.error(f"sandbox_client.py not found. Ensure the file exists.")
+
+        return f"#elitea simplified client\n{pyodide_predata}{code}"
+
+    def _initialize_sandbox(self) -> None:
+        """Initialize the PyodideSandbox instance with optimized settings"""
+        try:
+            # Check if Deno is available
+            if not _is_deno_available():
+                error_msg = (
+                    "Deno is required for PyodideSandbox but is not installed. "
+                    "Please run the bootstrap.sh script or install Deno manually."
+                )
+                logger.error(error_msg)
+                raise RuntimeError(error_msg)
+
+            from langchain_sandbox import PyodideSandbox
+
+            # Air-gapped settings
+            sandbox_base = os.environ.get("SANDBOX_BASE", os.path.expanduser('~/.cache/pyodide'))
+            sandbox_tmp = os.path.join(sandbox_base, "tmp")
+            deno_cache = os.environ.get("DENO_DIR", os.path.expanduser('~/.cache/deno'))
+
+            # Configure sandbox with performance optimizations
+            self._sandbox = PyodideSandbox(
+                stateful=self.stateful,
+                #
+                allow_env=["SANDBOX_BASE"],
+                allow_read=[sandbox_base, sandbox_tmp, deno_cache],
+                allow_write=[sandbox_tmp, deno_cache],
+                #
+                allow_net=self.allow_net,
+                # Use auto node_modules_dir for better caching
+                node_modules_dir="auto"
+            )
+            logger.info(f"PyodideSandbox initialized successfully (stateful={self.stateful})")
+        except ImportError as e:
+            if "langchain_sandbox" in str(e):
+                error_msg = (
+                    "langchain-sandbox is required for the PyodideSandboxTool. "
+                    "Please install it with: pip install langchain-sandbox"
+                )
+                logger.error(error_msg)
+                raise ImportError(error_msg) from e
+            else:
+                logger.error(f"Failed to import required module: {e}")
+                raise
+        except Exception as e:
+            logger.error(f"Failed to initialize PyodideSandbox: {e}")
+            raise
+
+    def _run(self, code: str) -> str:
+        """
+        Synchronous version - runs the async method in a new event loop
+        """
+        try:
+            # Check if sandbox is initialized, if not try to initialize
+            if self._sandbox is None:
+                self._initialize_sandbox()
+
+            # Prepare code with state and client injection
+            prepared_code = self._prepare_pyodide_input(code)
+
+            # Check if we're already in an async context
+            try:
+                loop = asyncio.get_running_loop()
+                # We're in an async context, but _run is supposed to be sync
+                # We'll need to use a different approach
+                import concurrent.futures
+                with concurrent.futures.ThreadPoolExecutor() as executor:
+                    future = executor.submit(asyncio.run, self._arun(prepared_code))
+                    return future.result()
+            except RuntimeError:
+                # No running loop, safe to use asyncio.run
+                return asyncio.run(self._arun(prepared_code))
+        except (ImportError, RuntimeError) as e:
+            # Handle specific dependency errors gracefully
+            error_msg = str(e)
+            if "langchain-sandbox" in error_msg:
+                return "❌ PyodideSandboxTool requires langchain-sandbox. Install with: pip install langchain-sandbox"
+            elif "Deno" in error_msg:
+                return "❌ PyodideSandboxTool requires Deno. Install from: https://docs.deno.com/runtime/getting_started/installation/"
+            else:
+                return f"❌ PyodideSandboxTool initialization failed: {error_msg}"
+        except Exception as e:
+            logger.error(f"Error executing code in sandbox: {e}")
+            return f"Error executing code: {str(e)}"
+
+    async def _arun(self, code: str) -> str:
+        """
+        Execute Python code in the Pyodide sandbox
+        """
+        try:
+            if self._sandbox is None:
+                self._initialize_sandbox()
+
+            # Execute the code with session state if available
+            result = await self._sandbox.execute(
+                code,
+                session_bytes=self.session_bytes,
+                session_metadata=self.session_metadata
+            )
+
+            # Update session state for stateful execution
+            if self.stateful:
+                self.session_bytes = result.session_bytes
+                self.session_metadata = result.session_metadata
+
+            result_dict = {}
+
+            if result.result is not None:
+                result_dict["result"] = result.result
+
+            if result.stdout:
+                result_dict["output"] = result.stdout
+
+            if result.stderr:
+                result_dict["error"] = result.stderr
+
+            if result.status == 'error':
+                result_dict["status"] = "Execution failed"
+
+            execution_info = f"Execution time: {result.execution_time:.2f}s"
+            if result.session_metadata and 'packages' in result.session_metadata:
+                packages = result.session_metadata.get('packages', [])
+                if packages:
+                    execution_info += f", Packages: {', '.join(packages)}"
+
+            result_dict["execution_info"] = execution_info
+            return result_dict
+
+        except Exception as e:
+            logger.error(f"Error executing code in sandbox: {e}")
+            return {"error": f"Error executing code: {str(e)}"}
+
+
+class StatefulPyodideSandboxTool(PyodideSandboxTool):
+    """
+    A stateful version of the PyodideSandboxTool that maintains state between executions.
+    This version preserves variables, imports, and function definitions across multiple tool calls.
+    """
+
+    name: str = "stateful_pyodide_sandbox"
+    description: str = """Execute Python code in a stateful sandbox environment using Pyodide.
+    This tool maintains state between executions, preserving variables, imports, and function definitions.
+    Use this tool when you need to:
+    - Build upon previous code executions
+    - Maintain variables across multiple calls
+    - Develop complex programs step by step
+    - Preserve imported libraries and defined functions
+
+    The sandbox supports most Python standard library modules and can install additional packages.
+    Note: File access and some system operations are restricted for security.
+    """
+
+    def __init__(self, **kwargs: Any) -> None:
+        kwargs['stateful'] = True  # Force stateful mode
+        super().__init__(**kwargs)
+
+
+# Factory function for creating sandbox tools
+def create_sandbox_tool(stateful: bool = False, allow_net: bool = True, alita_client: Optional[Any] = None) -> BaseTool:
+    """
+    Factory function to create sandbox tools with specified configuration.
+
+    Note: This tool requires Deno to be installed and available in PATH.
+    For installation and optimization, run the bootstrap.sh script.
+
+    Args:
+        stateful: Whether to maintain state between executions (default: False for better performance)
+        allow_net: Whether to allow network access (for package installation)
+
+    Returns:
+        Configured sandbox tool instance
+
+    Raises:
+        ImportError: If langchain-sandbox is not installed
+        RuntimeError: If Deno is not found in PATH
+
+    Performance Notes:
+        - Stateless mode (default) is faster and avoids session state overhead
+        - Run bootstrap.sh script to enable local caching and reduce initialization time
+        - Cached wheels reduce package download time from ~4.76s to near-instant
+    """
+    if stateful:
+        return StatefulPyodideSandboxTool(allow_net=allow_net, alita_client=alita_client)
+    else:
+        return PyodideSandboxTool(stateful=False, allow_net=allow_net, alita_client=alita_client)
+
+
+class SandboxToolkit(BaseToolkit):
+    tools: List[BaseTool] = []
+
+    @staticmethod
+    def toolkit_config_schema() -> Type[BaseModel]:
+        # Create sample tools to get their schemas
+        sample_tools = [
+            PyodideSandboxTool(),
+            StatefulPyodideSandboxTool()
+        ]
+        selected_tools = {x.name: x.args_schema.model_json_schema() for x in sample_tools}
+
+        return create_model(
+            'sandbox',
+            stateful=(bool, Field(default=False, description="Whether to maintain state between executions")),
+            allow_net=(bool, Field(default=True, description="Whether to allow network access for package installation")),
+            selected_tools=(List[Literal[tuple(selected_tools)]],
+                            Field(default=[], json_schema_extra={'args_schemas': selected_tools})),
+
+            __config__=ConfigDict(json_schema_extra={
+                'metadata': {
+                    "label": "Python Sandbox",
+                    "icon_url": "sandbox.svg",
+                    "hidden": False,
+                    "categories": ["code", "execution", "internal_tool"],
+                    "extra_categories": ["python", "pyodide", "sandbox", "code execution"],
+                }
+            })
+        )
+
+    @classmethod
+    def get_toolkit(cls, stateful: bool = False, allow_net: bool = True, alita_client=None, **kwargs):
+        """
+        Get toolkit with sandbox tools.
+
+        Args:
+            stateful: Whether to maintain state between executions
+            allow_net: Whether to allow network access
+            alita_client: Alita client instance for sandbox tools
+            **kwargs: Additional arguments
+        """
+        tools = []
+
+        if stateful:
+            tools.append(StatefulPyodideSandboxTool(allow_net=allow_net, alita_client=alita_client))
+        else:
+            tools.append(PyodideSandboxTool(stateful=False, allow_net=allow_net, alita_client=alita_client))
+
+        return cls(tools=tools)
+
+    def get_tools(self):
+        return self.tools

alita_sdk/runtime/tools/vectorstore.py

@@ -1,7 +1,7 @@
 import json
 import math
 import types
-from typing import Any, Optional, List, Dict, Callable, Generator
+from typing import Any, Optional, List, Dict, Callable, Generator, OrderedDict
 
 from langchain_core.documents import Document
 from pydantic import BaseModel, model_validator, Field
@@ -12,7 +12,6 @@ from alita_sdk.tools.vector_adapters.VectorStoreAdapter import VectorStoreAdapte
 from logging import getLogger
 
 from ..utils.logging import dispatch_custom_event
-from ..utils.utils import IndexerKeywords
 
 logger = getLogger(__name__)
 
@@ -73,6 +72,10 @@ class StepBackSearchDocumentsModel(BaseModel):
         }""",
         default=None
     )
+    extended_search: Optional[List[str]] = Field(
+        description="List of chunk types to search for (title, summary, propositions, keywords, documents)",
+        default=None
+    )
     reranking_config: Optional[Dict[str, Dict[str, Any]]] = Field(
         description="""Reranking configuration. Example:
         {
@@ -87,10 +90,6 @@ class StepBackSearchDocumentsModel(BaseModel):
         }""",
         default=None
     )
-    extended_search: Optional[List[str]] = Field(
-        description="List of chunk types to search for (title, summary, propositions, keywords, documents)",
-        default=None
-    )
 
 STEPBACK_PROMPT = """Your task is to convert provided question into a more generic question that will be used for similarity search.
 Remove all not important words, question words, but save all names, dates and acronym as in original question.
@@ -138,7 +137,7 @@ class VectorStoreWrapper(BaseToolApiWrapper):
     embedding_model_params: dict
     vectorstore_type: str
     vectorstore_params: dict
-    max_docs_per_add: int = 100
+    max_docs_per_add: int = 20
     dataset: str = None
     embedding: Any = None
     vectorstore: Any = None
@@ -208,16 +207,33 @@ class VectorStoreWrapper(BaseToolApiWrapper):
             tool_name="_remove_collection"
         )
 
-    def _get_indexed_ids(self, collection_suffix: Optional[str] = '') -> List[str]:
+    def _get_indexed_ids(self, index_name: Optional[str] = '') -> List[str]:
         """Get all indexed document IDs from vectorstore"""
-        return self.vector_adapter.get_indexed_ids(self, collection_suffix)
-
-    def list_collections(self) -> List[str]:
-        """List all collections in the vectorstore."""
-
-        return self.vector_adapter.list_collections(self)
+        return self.vector_adapter.get_indexed_ids(self, index_name)
+
+    def list_collections(self) -> Any:
+        """List all collections in the vectorstore.
+        Returns a list of collection names, or if no collections exist,
+        returns a dict with an empty list and a message."""
+        raw = self.vector_adapter.list_collections(self)
+        # Normalize raw result to a list of names
+        if not raw:
+            # No collections found
+            return {"collections": [], "message": "No indexed collections"}
+        if isinstance(raw, str):
+            # e.g., Chroma adapter returns comma-separated string
+            cols = [c for c in raw.split(',') if c]
+        else:
+            try:
+                cols = list(raw)
+            except Exception:
+                # Unexpected type, return raw directly
+                return raw
+        if not cols:
+            return {"collections": [], "message": "No indexed collections"}
+        return cols
 
-    def _clean_collection(self, collection_suffix: str = ''):
+    def _clean_collection(self, index_name: str = ''):
         """
         Clean the vectorstore collection by deleting all indexed data.
         """
@@ -225,19 +241,15 @@ class VectorStoreWrapper(BaseToolApiWrapper):
             f"Cleaning collection '{self.dataset}'",
             tool_name="_clean_collection"
         )
-        self.vector_adapter.clean_collection(self, collection_suffix)
+        self.vector_adapter.clean_collection(self, index_name)
         self._log_data(
             f"Collection '{self.dataset}' has been cleaned. ",
             tool_name="_clean_collection"
         )
 
-    def _get_indexed_data(self, collection_name: str):
-        """ Get all indexed data from vectorstore for non-code content """
-        return self.vector_adapter.get_indexed_data(self, collection_name)
-
-    def _get_code_indexed_data(self, collection_suffix: str) -> Dict[str, Dict[str, Any]]:
+    def _get_code_indexed_data(self, index_name: str) -> Dict[str, Dict[str, Any]]:
         """ Get all indexed data from vectorstore for code content """
-        return self.vector_adapter.get_code_indexed_data(self, collection_suffix)
+        return self.vector_adapter.get_code_indexed_data(self, index_name)
 
     def _add_to_collection(self, entry_id, new_collection_value):
         """Add a new collection name to the `collection` key in the `metadata` column."""
@@ -246,7 +258,7 @@ class VectorStoreWrapper(BaseToolApiWrapper):
     def _reduce_duplicates(
             self,
             documents: Generator[Any, None, None],
-            collection_suffix: str,
+            index_name: str,
             get_indexed_data: Callable,
             key_fn: Callable,
             compare_fn: Callable,
@@ -255,7 +267,7 @@ class VectorStoreWrapper(BaseToolApiWrapper):
     ) -> List[Any]:
         """Generic duplicate reduction logic for documents."""
         self._log_data(log_msg, tool_name="index_documents")
-        indexed_data = get_indexed_data(collection_suffix)
+        indexed_data = get_indexed_data(index_name)
         indexed_keys = set(indexed_data.keys())
         if not indexed_keys:
             self._log_data("Vectorstore is empty, indexing all incoming documents", tool_name="index_documents")
@@ -266,14 +278,15 @@ class VectorStoreWrapper(BaseToolApiWrapper):
 
         for document in documents:
             key = key_fn(document)
-            if key in indexed_keys and collection_suffix == indexed_data[key]['metadata'].get('collection'):
+            key = key if isinstance(key, str) else str(key)
+            if key in indexed_keys and index_name == indexed_data[key]['metadata'].get('collection'):
                 if compare_fn(document, indexed_data[key]):
                     # Disabled addition of new collection to already indexed documents
                     # # check metadata.collection and update if needed
                     # for update_collection_id in remove_ids_fn(indexed_data, key):
                     #     self._add_to_collection(
                     #         update_collection_id,
-                    #         collection_suffix
+                    #         index_name
                     #     )
                     continue
                 final_docs.append(document)
@@ -290,30 +303,10 @@ class VectorStoreWrapper(BaseToolApiWrapper):
 
         return final_docs
 
-    def _reduce_non_code_duplicates(self, documents: Generator[Any, None, None], collection_suffix: str) -> List[Any]:
-        return self._reduce_duplicates(
-            documents,
-            collection_suffix,
-            self._get_indexed_data,
-            lambda doc: doc.metadata.get('id'),
-            lambda doc, idx: (
-                    doc.metadata.get('updated_on') and
-                    idx['metadata'].get('updated_on') and
-                    doc.metadata.get('updated_on') == idx['metadata'].get('updated_on')
-            ),
-            lambda idx_data, key: (
-                    idx_data[key]['all_chunks'] +
-                    [idx_data[dep_id]['id'] for dep_id in idx_data[key][IndexerKeywords.DEPENDENT_DOCS.value]] +
-                    [chunk_db_id for dep_id in idx_data[key][IndexerKeywords.DEPENDENT_DOCS.value]
-                     for chunk_db_id in idx_data[dep_id]['all_chunks']]
-            ),
-            log_msg="Verification of documents to index started"
-        )
-
-    def _reduce_code_duplicates(self, documents: Generator[Any, None, None], collection_suffix: str) -> List[Any]:
+    def _reduce_code_duplicates(self, documents: Generator[Any, None, None], index_name: str) -> List[Any]:
         return self._reduce_duplicates(
             documents,
-            collection_suffix,
+            index_name,
             self._get_code_indexed_data,
             lambda doc: doc.metadata.get('filename'),
             lambda doc, idx: (
@@ -325,7 +318,7 @@ class VectorStoreWrapper(BaseToolApiWrapper):
             log_msg="Verification of code documents to index started"
         )
 
-    def index_documents(self, documents: Generator[Document, None, None], collection_suffix: str, progress_step: int = 20, clean_index: bool = True, is_code: bool = False):
+    def index_documents(self, documents: Generator[Document, None, None], index_name: str, progress_step: int = 20, clean_index: bool = True, is_code: bool = True):
         """ Index documents in the vectorstore.
 
         Args:
@@ -336,13 +329,13 @@ class VectorStoreWrapper(BaseToolApiWrapper):
 
         from ..langchain.interfaces.llm_processor import add_documents
 
-        self._log_tool_event(message=f"Starting the indexing... Parameters: {collection_suffix=}, {clean_index=}, {is_code}", tool_name="index_documents")
+        self._log_tool_event(message=f"Starting the indexing... Parameters: {index_name=}, {clean_index=}, {is_code}", tool_name="index_documents")
         # pre-process documents if needed (find duplicates, etc.)
         if clean_index:
             logger.info("Cleaning index before re-indexing all documents.")
             self._log_data("Cleaning index before re-indexing all documents. Previous index will be removed", tool_name="index_documents")
             try:
-                self._clean_collection(collection_suffix)
+                self._clean_collection(index_name)
                 self.vectoradapter.persist()
                 self.vectoradapter.vacuum()
                 self._log_data("Previous index has been removed",
@@ -356,8 +349,7 @@ class VectorStoreWrapper(BaseToolApiWrapper):
                 message="Filter for duplicates",
                 tool_name="index_documents")
             # remove duplicates based on metadata 'id' and 'updated_on' or 'commit_hash' fields
-            documents = self._reduce_code_duplicates(documents, collection_suffix) if is_code \
-                else self._reduce_non_code_duplicates(documents, collection_suffix)
+            documents = self._reduce_code_duplicates(documents, index_name)
             self._log_tool_event(
                 message="All the duplicates were filtered out. Proceeding with indexing.",
                 tool_name="index_documents")
@@ -385,13 +377,13 @@ class VectorStoreWrapper(BaseToolApiWrapper):
         self._log_tool_event(message=f"Documents for indexing were processed. Total documents: {len(documents)}",
                              tool_name="index_documents")
 
-        # if collection_suffix is provided, add it to metadata of each document
-        if collection_suffix:
+        # if index_name is provided, add it to metadata of each document
+        if index_name:
             for doc in documents:
                 if not doc.metadata.get('collection'):
-                    doc.metadata['collection'] = collection_suffix
+                    doc.metadata['collection'] = index_name
                 else:
-                    doc.metadata['collection'] += f";{collection_suffix}"
+                    doc.metadata['collection'] += f";{index_name}"
 
         total_docs = len(documents)
         documents_count = 0
@@ -422,7 +414,8 @@ class VectorStoreWrapper(BaseToolApiWrapper):
                 return {"status": "error", "message": f"Error: {format_exc()}"}
         if _documents:
             add_documents(vectorstore=self.vectorstore, documents=_documents)
-        return {"status": "ok", "message": f"successfully indexed {documents_count} documents"}
+        return {"status": "ok", "message": f"successfully indexed {documents_count} documents" if documents_count > 0
+        else "No new documents to index."}
 
     def search_documents(self, query:str, doctype: str = 'code', 
                          filter:dict|str={}, cut_off: float=0.5,
@@ -542,11 +535,18 @@ class VectorStoreWrapper(BaseToolApiWrapper):
             
         # Initialize document map for tracking by ID
         doc_map = {
-            f"{doc.metadata.get('id', f'idx_{i}')}_{doc.metadata['chunk_id']}"
-            if 'chunk_id' in doc.metadata
-            else doc.metadata.get('id', f"idx_{i}"): (doc, score)
+            (
+                f"{doc.metadata.get('id', f'idx_{i}')}_{doc.metadata['chunk_id']}"
+                if 'chunk_id' in doc.metadata
+                else doc.metadata.get('id', f"idx_{i}")
+            ): (doc, 1 - score)
             for i, (doc, score) in enumerate(vector_items)
         }
+
+        # Sort the items by the new score in descending order
+        doc_map = OrderedDict(
+            sorted(doc_map.items(), key=lambda x: x[1][1], reverse=True)
+        )
         
         # Process full-text search if configured
         if full_text_search and full_text_search.get('enabled') and full_text_search.get('fields'):
@@ -597,7 +597,7 @@ class VectorStoreWrapper(BaseToolApiWrapper):
         # Apply cutoff filter
         if cut_off:
             # Filter out items above the cutoff score (since the lower the score, the better)
-            combined_items = [item for item in combined_items if abs(item[1]) <= cut_off]
+            combined_items = [item for item in combined_items if abs(item[1]) >= cut_off]
         
         # Sort by score and limit results
         # DISABLED: for chroma we want ascending order (lower score is better), for others descending
@@ -758,4 +758,3 @@ class VectorStoreWrapper(BaseToolApiWrapper):
                 "args_schema": StepBackSearchDocumentsModel
             }
         ]
-