mcp-vector-search 0.12.6__py3-none-any.whl → 1.0.3__py3-none-any.whl

mcp_vector_search/cli/commands/chat.py

@@ -0,0 +1,1262 @@
+"""Chat command for LLM-powered intelligent code search."""
+
+import asyncio
+import os
+from fnmatch import fnmatch
+from pathlib import Path
+from typing import Any
+
+import typer
+from loguru import logger
+from rich.live import Live
+from rich.markdown import Markdown
+from rich.panel import Panel
+
+from ...core.database import ChromaVectorDatabase
+from ...core.embeddings import create_embedding_function
+from ...core.exceptions import ProjectNotFoundError, SearchError
+from ...core.llm_client import LLMClient
+from ...core.project import ProjectManager
+from ...core.search import SemanticSearchEngine
+from ..didyoumean import create_enhanced_typer
+from ..output import (
+    console,
+    print_error,
+    print_info,
+    print_success,
+    print_warning,
+)
+
+
+def show_api_key_help() -> None:
+    """Display helpful error message when API key is missing."""
+    message = """[bold yellow]⚠️  No LLM API Key Found[/bold yellow]
+
+The chat feature requires an API key for an LLM provider.
+
+[bold cyan]Set one of these environment variables:[/bold cyan]
+  • [green]OPENAI_API_KEY[/green]       - For OpenAI (GPT-4, etc.) [dim](recommended)[/dim]
+  • [green]OPENROUTER_API_KEY[/green]  - For OpenRouter (Claude, GPT, etc.)
+
+[bold cyan]Example:[/bold cyan]
+  [yellow]export OPENAI_API_KEY="sk-..."[/yellow]
+  [yellow]export OPENROUTER_API_KEY="sk-or-..."[/yellow]
+
+[bold cyan]Get API keys at:[/bold cyan]
+  • OpenAI: [link=https://platform.openai.com/api-keys]https://platform.openai.com/api-keys[/link]
+  • OpenRouter: [link=https://openrouter.ai/keys]https://openrouter.ai/keys[/link]
+
+[dim]Alternatively, run: [cyan]mcp-vector-search setup[/cyan] for interactive setup[/dim]"""
+
+    panel = Panel(
+        message,
+        border_style="yellow",
+        padding=(1, 2),
+    )
+    console.print(panel)
+
+
+class ChatSession:
+    """Manages conversation history with automatic compaction.
+
+    Keeps system prompt intact, compacts older messages when history grows large,
+    and maintains recent exchanges for context.
+    """
+
+    # Threshold for compaction (estimated tokens, ~4 chars per token)
+    COMPACTION_THRESHOLD = 8000 * 4  # ~32000 chars
+    RECENT_EXCHANGES_TO_KEEP = 3  # Keep last N user/assistant pairs
+
+    def __init__(self, system_prompt: str) -> None:
+        """Initialize session with system prompt.
+
+        Args:
+            system_prompt: Initial system message
+        """
+        self.system_prompt = system_prompt
+        self.messages: list[dict[str, str]] = [
+            {"role": "system", "content": system_prompt}
+        ]
+
+    def add_message(self, role: str, content: str) -> None:
+        """Add message to history and compact if needed.
+
+        Args:
+            role: Message role (user/assistant)
+            content: Message content
+        """
+        self.messages.append({"role": role, "content": content})
+
+        # Check if compaction needed
+        total_chars = sum(len(msg["content"]) for msg in self.messages)
+        if total_chars > self.COMPACTION_THRESHOLD:
+            self._compact_history()
+
+    def _compact_history(self) -> None:
+        """Compact conversation history by summarizing older exchanges.
+
+        Strategy:
+        1. Keep system prompt intact
+        2. Summarize older exchanges into brief context
+        3. Keep recent N exchanges verbatim
+        """
+        logger.debug("Compacting conversation history")
+
+        # Separate system prompt and conversation
+        system_msg = self.messages[0]
+        conversation = self.messages[1:]
+
+        # Keep recent exchanges (last N user/assistant pairs)
+        recent_start = max(0, len(conversation) - (self.RECENT_EXCHANGES_TO_KEEP * 2))
+        older_messages = conversation[:recent_start]
+        recent_messages = conversation[recent_start:]
+
+        # Summarize older messages
+        if older_messages:
+            summary_parts = []
+            for msg in older_messages:
+                role = msg["role"].capitalize()
+                content_preview = msg["content"][:100].replace("\n", " ")
+                summary_parts.append(f"{role}: {content_preview}...")
+
+            summary = "\n".join(summary_parts)
+            summary_msg = {
+                "role": "system",
+                "content": f"[Previous conversation summary]\n{summary}\n[End summary]",
+            }
+
+            # Rebuild messages: system + summary + recent
+            self.messages = [system_msg, summary_msg] + recent_messages
+
+            logger.debug(
+                f"Compacted {len(older_messages)} old messages, kept {len(recent_messages)} recent"
+            )
+
+    def get_messages(self) -> list[dict[str, str]]:
+        """Get current message history.
+
+        Returns:
+            List of message dictionaries
+        """
+        return self.messages.copy()
+
+    def clear(self) -> None:
+        """Clear conversation history, keeping only system prompt."""
+        self.messages = [{"role": "system", "content": self.system_prompt}]
+
+
+# Create chat subcommand app with "did you mean" functionality
+chat_app = create_enhanced_typer(
+    help="🤖 LLM-powered intelligent code search",
+    invoke_without_command=True,
+)
+
+
+@chat_app.callback(invoke_without_command=True)
+def chat_main(
+    ctx: typer.Context,
+    query: str | None = typer.Argument(
+        None,
+        help="Natural language query about your code",
+    ),
+    project_root: Path | None = typer.Option(
+        None,
+        "--project-root",
+        "-p",
+        help="Project root directory (auto-detected if not specified)",
+        exists=True,
+        file_okay=False,
+        dir_okay=True,
+        readable=True,
+        rich_help_panel="🔧 Global Options",
+    ),
+    limit: int = typer.Option(
+        5,
+        "--limit",
+        "-l",
+        help="Maximum number of results to return",
+        min=1,
+        max=20,
+        rich_help_panel="📊 Result Options",
+    ),
+    model: str | None = typer.Option(
+        None,
+        "--model",
+        "-m",
+        help="Model to use (defaults based on provider: gpt-4o-mini for OpenAI, claude-3-haiku for OpenRouter)",
+        rich_help_panel="🤖 LLM Options",
+    ),
+    provider: str | None = typer.Option(
+        None,
+        "--provider",
+        help="LLM provider to use: 'openai' or 'openrouter' (auto-detect if not specified)",
+        rich_help_panel="🤖 LLM Options",
+    ),
+    timeout: float | None = typer.Option(
+        30.0,
+        "--timeout",
+        help="API timeout in seconds",
+        min=5.0,
+        max=120.0,
+        rich_help_panel="🤖 LLM Options",
+    ),
+    json_output: bool = typer.Option(
+        False,
+        "--json",
+        help="Output results in JSON format",
+        rich_help_panel="📊 Result Options",
+    ),
+    files: str | None = typer.Option(
+        None,
+        "--files",
+        "-f",
+        help="Filter by file glob patterns (e.g., '*.py', 'src/*.js'). Matches basename or relative path.",
+        rich_help_panel="🔍 Filters",
+    ),
+    think: bool = typer.Option(
+        False,
+        "--think",
+        "-t",
+        help="Use advanced model for complex queries (gpt-4o / claude-sonnet-4). Better reasoning, higher cost.",
+        rich_help_panel="🤖 LLM Options",
+    ),
+) -> None:
+    """🤖 Ask questions about your code in natural language.
+
+    Uses LLM (OpenAI or OpenRouter) to intelligently search your codebase and answer
+    questions like "where is X defined?", "how does Y work?", etc.
+
+    [bold cyan]Setup:[/bold cyan]
+
+    [green]Option A - OpenAI (recommended):[/green]
+        $ export OPENAI_API_KEY="your-key-here"
+        Get a key at: [cyan]https://platform.openai.com/api-keys[/cyan]
+
+    [green]Option B - OpenRouter:[/green]
+        $ export OPENROUTER_API_KEY="your-key-here"
+        Get a key at: [cyan]https://openrouter.ai/keys[/cyan]
+
+    [dim]Provider is auto-detected. OpenAI is preferred if both keys are set.[/dim]
+
+    [bold cyan]Examples:[/bold cyan]
+
+    [green]Ask where a parameter is set:[/green]
+        $ mcp-vector-search chat "where is similarity_threshold set?"
+
+    [green]Ask how something works:[/green]
+        $ mcp-vector-search chat "how does the indexing process work?"
+
+    [green]Find implementation details:[/green]
+        $ mcp-vector-search chat "show me the search ranking algorithm"
+
+    [green]Force specific provider:[/green]
+        $ mcp-vector-search chat "question" --provider openai
+        $ mcp-vector-search chat "question" --provider openrouter
+
+    [green]Use custom model:[/green]
+        $ mcp-vector-search chat "question" --model gpt-4o
+        $ mcp-vector-search chat "question" --model anthropic/claude-3.5-sonnet
+
+    [bold cyan]Advanced:[/bold cyan]
+
+    [green]Filter by file pattern:[/green]
+        $ mcp-vector-search chat "how does validation work?" --files "*.py"
+        $ mcp-vector-search chat "find React components" --files "src/*.tsx"
+
+    [green]Limit results:[/green]
+        $ mcp-vector-search chat "find auth code" --limit 3
+
+    [green]Custom timeout:[/green]
+        $ mcp-vector-search chat "complex question" --timeout 60
+
+    [green]Use advanced model for complex queries:[/green]
+        $ mcp-vector-search chat "explain the authentication flow" --think
+
+    [dim]💡 Tip: Use --think for complex architectural questions. It uses gpt-4o or
+    claude-sonnet-4 for better reasoning at higher cost.[/dim]
+    """
+    # If no query provided and no subcommand invoked, exit (show help)
+    if query is None:
+        if ctx.invoked_subcommand is None:
+            # No query and no subcommand - show help
+            raise typer.Exit()
+        else:
+            # A subcommand was invoked - let it handle the request
+            return
+
+    try:
+        project_root = project_root or ctx.obj.get("project_root") or Path.cwd()
+
+        # Validate provider if specified
+        if provider and provider not in ("openai", "openrouter"):
+            print_error(
+                f"Invalid provider: {provider}. Must be 'openai' or 'openrouter'"
+            )
+            raise typer.Exit(1)
+
+        # Run the chat with intent detection and routing
+        asyncio.run(
+            run_chat_with_intent(
+                project_root=project_root,
+                query=query,
+                limit=limit,
+                model=model,
+                provider=provider,
+                timeout=timeout,
+                json_output=json_output,
+                files=files,
+                think=think,
+            )
+        )
+
+    except (typer.Exit, SystemExit):
+        # Re-raise exit exceptions without printing additional error messages
+        # The error message has already been shown to the user
+        raise
+    except Exception as e:
+        #  Log real exceptions (not typer.Exit)
+        if not isinstance(e, typer.Exit | SystemExit):
+            logger.error(f"Chat failed: {e}")
+            print_error(f"Chat failed: {e}")
+        raise typer.Exit(
+            1
+        ) from None  # Suppress exception chain to avoid double-printing
+
+
+async def run_chat_with_intent(
+    project_root: Path,
+    query: str,
+    limit: int = 5,
+    model: str | None = None,
+    provider: str | None = None,
+    timeout: float = 30.0,
+    json_output: bool = False,
+    files: str | None = None,
+    think: bool = False,
+) -> None:
+    """Route to appropriate chat mode based on detected intent.
+
+    Args:
+        project_root: Project root directory
+        query: User's natural language query
+        limit: Maximum results to return
+        model: Model to use (optional)
+        provider: LLM provider
+        timeout: API timeout
+        json_output: Whether to output JSON
+        files: File pattern filter
+        think: Use advanced model
+    """
+    # Initialize LLM client for intent detection
+    from ...core.config_utils import (
+        get_openai_api_key,
+        get_openrouter_api_key,
+        get_preferred_llm_provider,
+    )
+
+    config_dir = project_root / ".mcp-vector-search"
+    openai_key = get_openai_api_key(config_dir)
+    openrouter_key = get_openrouter_api_key(config_dir)
+
+    # Determine provider (same logic as before)
+    if not provider:
+        preferred_provider = get_preferred_llm_provider(config_dir)
+        if preferred_provider == "openai" and openai_key:
+            provider = "openai"
+        elif preferred_provider == "openrouter" and openrouter_key:
+            provider = "openrouter"
+        elif openai_key:
+            provider = "openai"
+        elif openrouter_key:
+            provider = "openrouter"
+        else:
+            console.print()  # Blank line for spacing
+            show_api_key_help()
+            raise typer.Exit(1)
+
+    # Create temporary client for intent detection (use fast model)
+    try:
+        intent_client = LLMClient(
+            openai_api_key=openai_key,
+            openrouter_api_key=openrouter_key,
+            provider=provider,
+            timeout=timeout,
+            think=False,  # Use fast model for intent detection
+        )
+
+        # Detect intent
+        intent = await intent_client.detect_intent(query)
+
+        # Show intent to user
+        if intent == "find":
+            console.print("\n[cyan]🔍 Intent: Find[/cyan] - Searching codebase\n")
+            await run_chat_search(
+                project_root=project_root,
+                query=query,
+                limit=limit,
+                model=model,
+                provider=provider,
+                timeout=timeout,
+                json_output=json_output,
+                files=files,
+                think=think,
+            )
+        else:
+            # Answer mode - force think mode and enter interactive session
+            console.print(
+                "\n[cyan]💬 Intent: Answer[/cyan] - Entering interactive mode\n"
+            )
+            await run_chat_answer(
+                project_root=project_root,
+                initial_query=query,
+                limit=limit,
+                model=model,
+                provider=provider,
+                timeout=timeout,
+                files=files,
+            )
+
+    except Exception as e:
+        logger.error(f"Intent detection failed: {e}")
+        # Default to find mode on error
+        console.print("\n[yellow]⚠ Using default search mode[/yellow]\n")
+        await run_chat_search(
+            project_root=project_root,
+            query=query,
+            limit=limit,
+            model=model,
+            provider=provider,
+            timeout=timeout,
+            json_output=json_output,
+            files=files,
+            think=think,
+        )
+
+
+async def run_chat_answer(
+    project_root: Path,
+    initial_query: str,
+    limit: int = 5,
+    model: str | None = None,
+    provider: str | None = None,
+    timeout: float = 30.0,
+    files: str | None = None,
+) -> None:
+    """Run interactive answer mode with streaming responses.
+
+    Args:
+        project_root: Project root directory
+        initial_query: Initial user question
+        limit: Max search results for context
+        model: Model to use (optional, defaults to advanced model)
+        provider: LLM provider
+        timeout: API timeout
+        files: File pattern filter
+    """
+    from ...core.config_utils import get_openai_api_key, get_openrouter_api_key
+
+    config_dir = project_root / ".mcp-vector-search"
+    openai_key = get_openai_api_key(config_dir)
+    openrouter_key = get_openrouter_api_key(config_dir)
+
+    # Load project configuration
+    project_manager = ProjectManager(project_root)
+    if not project_manager.is_initialized():
+        raise ProjectNotFoundError(
+            f"Project not initialized at {project_root}. Run 'mcp-vector-search init' first."
+        )
+
+    config = project_manager.load_config()
+
+    # Initialize LLM client with advanced model (force think mode)
+    try:
+        llm_client = LLMClient(
+            openai_api_key=openai_key,
+            openrouter_api_key=openrouter_key,
+            model=model,
+            provider=provider,
+            timeout=timeout,
+            think=True,  # Always use advanced model for answer mode
+        )
+        provider_display = llm_client.provider.capitalize()
+        model_info = f"{llm_client.model} [bold magenta](thinking mode)[/bold magenta]"
+        print_success(f"Connected to {provider_display}: {model_info}")
+    except ValueError as e:
+        print_error(str(e))
+        raise typer.Exit(1)
+
+    # Initialize search engine
+    embedding_function, _ = create_embedding_function(config.embedding_model)
+    database = ChromaVectorDatabase(
+        persist_directory=config.index_path,
+        embedding_function=embedding_function,
+    )
+    search_engine = SemanticSearchEngine(
+        database=database,
+        project_root=project_root,
+        similarity_threshold=config.similarity_threshold,
+    )
+
+    # Initialize session (cleared on startup)
+    system_prompt = """You are a helpful code assistant analyzing a codebase. Answer questions based on provided code context.
+
+Guidelines:
+- Be concise but thorough
+- Reference specific functions, classes, or files
+- Use code examples when helpful
+- If context is insufficient, say so
+- Use markdown formatting"""
+
+    session = ChatSession(system_prompt)
+
+    # Process initial query
+    await _process_answer_query(
+        query=initial_query,
+        llm_client=llm_client,
+        search_engine=search_engine,
+        database=database,
+        session=session,
+        project_root=project_root,
+        limit=limit,
+        files=files,
+        config=config,
+    )
+
+    # Interactive loop
+    console.print("\n[dim]Type your questions or '/exit' to quit[/dim]\n")
+
+    while True:
+        try:
+            # Get user input
+            user_input = console.input("\n[bold cyan]You:[/bold cyan] ").strip()
+
+            if not user_input:
+                continue
+
+            # Check for exit command
+            if user_input.lower() in ("/exit", "/quit", "exit", "quit"):
+                console.print("\n[cyan]👋 Session ended.[/cyan]")
+                break
+
+            # Process query
+            await _process_answer_query(
+                query=user_input,
+                llm_client=llm_client,
+                search_engine=search_engine,
+                database=database,
+                session=session,
+                project_root=project_root,
+                limit=limit,
+                files=files,
+                config=config,
+            )
+
+        except KeyboardInterrupt:
+            console.print("\n\n[cyan]👋 Session ended.[/cyan]")
+            break
+        except EOFError:
+            console.print("\n\n[cyan]👋 Session ended.[/cyan]")
+            break
+        except Exception as e:
+            logger.error(f"Error processing query: {e}")
+            print_error(f"Error: {e}")
+
+
+async def _process_answer_query(
+    query: str,
+    llm_client: LLMClient,
+    search_engine: SemanticSearchEngine,
+    database: ChromaVectorDatabase,
+    session: ChatSession,
+    project_root: Path,
+    limit: int,
+    files: str | None,
+    config: Any,
+) -> None:
+    """Process a single answer query with agentic tool use.
+
+    Args:
+        query: User query
+        llm_client: LLM client instance
+        search_engine: Search engine instance
+        database: Vector database
+        session: Chat session
+        project_root: Project root path
+        limit: Max results
+        files: File pattern filter
+        config: Project config
+    """
+    # Define search tools for the LLM
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "search_code",
+                "description": "Search the codebase for relevant code snippets using semantic search",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "query": {
+                            "type": "string",
+                            "description": "Search query to find relevant code (e.g., 'authentication logic', 'database connection', 'error handling')",
+                        },
+                        "limit": {
+                            "type": "integer",
+                            "description": "Maximum number of results to return (default: 5, max: 10)",
+                            "default": 5,
+                        },
+                    },
+                    "required": ["query"],
+                },
+            },
+        },
+        {
+            "type": "function",
+            "function": {
+                "name": "read_file",
+                "description": "Read the full content of a specific file",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "file_path": {
+                            "type": "string",
+                            "description": "Relative path to the file from project root",
+                        }
+                    },
+                    "required": ["file_path"],
+                },
+            },
+        },
+        {
+            "type": "function",
+            "function": {
+                "name": "list_files",
+                "description": "List files in the codebase matching a pattern",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "pattern": {
+                            "type": "string",
+                            "description": "Glob pattern to match files (e.g., '*.py', 'src/**/*.ts', 'tests/')",
+                        }
+                    },
+                    "required": ["pattern"],
+                },
+            },
+        },
+    ]
+
+    # System prompt for tool use
+    system_prompt = """You are a helpful code assistant with access to search tools. Use these tools to find and analyze code in the codebase.
+
+Available tools:
+- search_code: Search for relevant code using semantic search
+- read_file: Read the full content of a specific file
+- list_files: List files matching a pattern
+
+Guidelines:
+1. Use search_code to find relevant code snippets
+2. Use read_file when you need to see the full file context
+3. Use list_files to understand the project structure
+4. Make multiple searches if needed to gather enough context
+5. After gathering sufficient information, provide your analysis
+
+Always base your answers on actual code from the tools. If you can't find relevant code, say so."""
+
+    # Tool execution functions
+    async def execute_search_code(query_str: str, limit_val: int = 5) -> str:
+        """Execute search_code tool."""
+        try:
+            limit_val = min(limit_val, 10)  # Cap at 10
+            async with database:
+                results = await search_engine.search(
+                    query=query_str,
+                    limit=limit_val,
+                    similarity_threshold=config.similarity_threshold,
+                    include_context=True,
+                )
+
+                # Post-filter by file pattern if specified
+                if files and results:
+                    filtered_results = []
+                    for result in results:
+                        try:
+                            rel_path = str(result.file_path.relative_to(project_root))
+                        except ValueError:
+                            rel_path = str(result.file_path)
+
+                        if fnmatch(rel_path, files) or fnmatch(
+                            os.path.basename(rel_path), files
+                        ):
+                            filtered_results.append(result)
+                    results = filtered_results
+
+            if not results:
+                return "No results found for this query."
+
+            # Format results
+            result_parts = []
+            for i, result in enumerate(results, 1):
+                try:
+                    rel_path = str(result.file_path.relative_to(project_root))
+                except ValueError:
+                    rel_path = str(result.file_path)
+
+                result_parts.append(
+                    f"[Result {i}: {rel_path}]\n"
+                    f"Location: {result.location}\n"
+                    f"Lines {result.start_line}-{result.end_line}\n"
+                    f"Similarity: {result.similarity_score:.3f}\n"
+                    f"```\n{result.content}\n```\n"
+                )
+            return "\n".join(result_parts)
+
+        except Exception as e:
+            logger.error(f"search_code tool failed: {e}")
+            return f"Error searching code: {e}"
+
+    async def execute_read_file(file_path: str) -> str:
+        """Execute read_file tool."""
+        try:
+            # Normalize path
+            if file_path.startswith("/"):
+                full_path = Path(file_path)
+            else:
+                full_path = project_root / file_path
+
+            # Security check: file must be within project
+            try:
+                full_path.relative_to(project_root)
+            except ValueError:
+                return f"Error: File must be within project root: {project_root}"
+
+            if not full_path.exists():
+                return f"Error: File not found: {file_path}"
+
+            if not full_path.is_file():
+                return f"Error: Not a file: {file_path}"
+
+            # Read file with size limit
+            max_size = 100_000  # 100KB
+            file_size = full_path.stat().st_size
+            if file_size > max_size:
+                return f"Error: File too large ({file_size} bytes). Use search_code instead."
+
+            content = full_path.read_text(errors="replace")
+            return f"File: {file_path}\n```\n{content}\n```"
+
+        except Exception as e:
+            logger.error(f"read_file tool failed: {e}")
+            return f"Error reading file: {e}"
+
+    async def execute_list_files(pattern: str) -> str:
+        """Execute list_files tool."""
+        try:
+            from glob import glob
+
+            # Use glob to find matching files
+            matches = glob(str(project_root / pattern), recursive=True)
+
+            if not matches:
+                return f"No files found matching pattern: {pattern}"
+
+            # Get relative paths and limit results
+            rel_paths = []
+            for match in matches[:50]:  # Limit to 50 files
+                try:
+                    rel_path = Path(match).relative_to(project_root)
+                    rel_paths.append(str(rel_path))
+                except ValueError:
+                    continue
+
+            if not rel_paths:
+                return f"No files found matching pattern: {pattern}"
+
+            return f"Files matching '{pattern}':\n" + "\n".join(
+                f"- {p}" for p in sorted(rel_paths)
+            )
+
+        except Exception as e:
+            logger.error(f"list_files tool failed: {e}")
+            return f"Error listing files: {e}"
+
+    # Get conversation history
+    conversation_history = session.get_messages()[1:]  # Skip system prompt
+
+    # Build messages: system + history + current query
+    messages = [{"role": "system", "content": system_prompt}]
+    messages.extend(conversation_history)
+    messages.append({"role": "user", "content": query})
+
+    # Agentic loop
+    max_iterations = 25
+    for _iteration in range(max_iterations):
+        try:
+            response = await llm_client.chat_with_tools(messages, tools)
+
+            # Extract message from response
+            choice = response.get("choices", [{}])[0]
+            message = choice.get("message", {})
+
+            # Check for tool calls
+            tool_calls = message.get("tool_calls", [])
+
+            if tool_calls:
+                # Add assistant message with tool calls
+                messages.append(message)
+
+                # Execute each tool call
+                for tool_call in tool_calls:
+                    tool_id = tool_call.get("id")
+                    function = tool_call.get("function", {})
+                    function_name = function.get("name")
+                    arguments_str = function.get("arguments", "{}")
+
+                    # Parse arguments
+                    try:
+                        import json
+
+                        arguments = json.loads(arguments_str)
+                    except json.JSONDecodeError:
+                        arguments = {}
+
+                    # Display tool usage
+                    console.print(
+                        f"\n[dim]🔧 Using tool: {function_name}({', '.join(f'{k}={repr(v)}' for k, v in arguments.items())})[/dim]"
+                    )
+
+                    # Execute tool
+                    if function_name == "search_code":
+                        result = await execute_search_code(
+                            arguments.get("query", ""),
+                            arguments.get("limit", 5),
+                        )
+                        console.print(
+                            f"[dim]   Found {len(result.split('[Result')) - 1} results[/dim]"
+                        )
+                    elif function_name == "read_file":
+                        result = await execute_read_file(arguments.get("file_path", ""))
+                        console.print("[dim]   Read file[/dim]")
+                    elif function_name == "list_files":
+                        result = await execute_list_files(arguments.get("pattern", ""))
+                        console.print("[dim]   Listed files[/dim]")
+                    else:
+                        result = f"Error: Unknown tool: {function_name}"
+
+                    # Add tool result to messages
+                    messages.append(
+                        {
+                            "role": "tool",
+                            "tool_call_id": tool_id,
+                            "content": result,
+                        }
+                    )
+
+            else:
+                # No tool calls - final response
+                final_content = message.get("content", "")
+
+                if not final_content:
+                    print_error("LLM returned empty response")
+                    return
+
+                # Stream the final response
+                console.print("\n[bold cyan]🤖 Assistant:[/bold cyan]\n")
+
+                # Use Rich Live for rendering
+                with Live("", console=console, auto_refresh=True) as live:
+                    live.update(Markdown(final_content))
+
+                # Add to session history
+                session.add_message("user", query)
+                session.add_message("assistant", final_content)
+
+                return
+
+        except Exception as e:
+            logger.error(f"Tool execution loop failed: {e}")
+            print_error(f"Error: {e}")
+            return
+
+    # Max iterations reached
+    print_warning(
+        "\n⚠ Maximum iterations reached. The assistant may not have gathered enough information."
+    )
+
+
+async def run_chat_search(
+    project_root: Path,
+    query: str,
+    limit: int = 5,
+    model: str | None = None,
+    provider: str | None = None,
+    timeout: float = 30.0,
+    json_output: bool = False,
+    files: str | None = None,
+    think: bool = False,
+) -> None:
+    """Run LLM-powered chat search.
+
+    Implementation Flow:
+    1. Initialize LLM client and validate API key
+    2. Generate 2-3 targeted search queries from natural language
+    3. Execute each search query against vector database
+    4. Have LLM analyze all results and select most relevant ones
+    5. Display results with explanations
+
+    Args:
+        project_root: Project root directory
+        query: Natural language query from user
+        limit: Maximum number of results to return
+        model: Model to use (optional, defaults based on provider)
+        provider: LLM provider ('openai' or 'openrouter', auto-detect if None)
+        timeout: API timeout in seconds
+        json_output: Whether to output JSON format
+        files: Optional glob pattern to filter files (e.g., '*.py', 'src/*.js')
+        think: Use advanced "thinking" model for complex queries
+    """
+    # Check for API keys (environment variable or config file)
+    from ...core.config_utils import (
+        get_openai_api_key,
+        get_openrouter_api_key,
+        get_preferred_llm_provider,
+    )
+
+    config_dir = project_root / ".mcp-vector-search"
+    openai_key = get_openai_api_key(config_dir)
+    openrouter_key = get_openrouter_api_key(config_dir)
+
+    # Determine which provider to use
+    if provider:
+        # Explicit provider specified
+        if provider == "openai" and not openai_key:
+            print_error("OpenAI API key not found.")
+            print_info("\n[bold]To use OpenAI:[/bold]")
+            print_info(
+                "1. Get an API key from [cyan]https://platform.openai.com/api-keys[/cyan]"
+            )
+            print_info("2. Set environment variable:")
+            print_info("   [yellow]export OPENAI_API_KEY='your-key'[/yellow]")
+            print_info("")
+            print_info("Or run: [cyan]mcp-vector-search setup[/cyan]")
+            raise typer.Exit(1)
+        elif provider == "openrouter" and not openrouter_key:
+            print_error("OpenRouter API key not found.")
+            print_info("\n[bold]To use OpenRouter:[/bold]")
+            print_info("1. Get an API key from [cyan]https://openrouter.ai/keys[/cyan]")
+            print_info("2. Set environment variable:")
+            print_info("   [yellow]export OPENROUTER_API_KEY='your-key'[/yellow]")
+            print_info("")
+            print_info("Or run: [cyan]mcp-vector-search setup[/cyan]")
+            raise typer.Exit(1)
+    else:
+        # Auto-detect provider
+        preferred_provider = get_preferred_llm_provider(config_dir)
+
+        if preferred_provider == "openai" and openai_key:
+            provider = "openai"
+        elif preferred_provider == "openrouter" and openrouter_key:
+            provider = "openrouter"
+        elif openai_key:
+            provider = "openai"
+        elif openrouter_key:
+            provider = "openrouter"
+        else:
+            console.print()  # Blank line for spacing
+            show_api_key_help()
+            raise typer.Exit(1)
+
+    # Load project configuration
+    project_manager = ProjectManager(project_root)
+
+    if not project_manager.is_initialized():
+        raise ProjectNotFoundError(
+            f"Project not initialized at {project_root}. Run 'mcp-vector-search init' first."
+        )
+
+    config = project_manager.load_config()
+
+    # Initialize LLM client
+    try:
+        llm_client = LLMClient(
+            openai_api_key=openai_key,
+            openrouter_api_key=openrouter_key,
+            model=model,
+            provider=provider,
+            timeout=timeout,
+            think=think,
+        )
+        provider_display = llm_client.provider.capitalize()
+        model_info = f"{llm_client.model}"
+        if think:
+            model_info += " [bold magenta](thinking mode)[/bold magenta]"
+        print_success(f"Connected to {provider_display}: {model_info}")
+    except ValueError as e:
+        print_error(str(e))
+        raise typer.Exit(1)
+
+    # Step 1: Generate search queries from natural language
+    console.print(f"\n[cyan]💭 Analyzing query:[/cyan] [white]{query}[/white]")
+
+    try:
+        search_queries = await llm_client.generate_search_queries(query, limit=3)
+
+        if not search_queries:
+            print_error("Failed to generate search queries from your question.")
+            raise typer.Exit(1)
+
+        console.print(
+            f"\n[cyan]🔍 Generated {len(search_queries)} search queries:[/cyan]"
+        )
+        for i, sq in enumerate(search_queries, 1):
+            console.print(f"  {i}. [yellow]{sq}[/yellow]")
+
+    except SearchError as e:
+        print_error(f"Failed to generate queries: {e}")
+        raise typer.Exit(1)
+
+    # Step 2: Execute each search query
+    console.print("\n[cyan]🔎 Searching codebase...[/cyan]")
+
+    embedding_function, _ = create_embedding_function(config.embedding_model)
+    database = ChromaVectorDatabase(
+        persist_directory=config.index_path,
+        embedding_function=embedding_function,
+    )
+
+    search_engine = SemanticSearchEngine(
+        database=database,
+        project_root=project_root,
+        similarity_threshold=config.similarity_threshold,
+    )
+
+    # Execute all searches
+    search_results = {}
+    total_results = 0
+
+    try:
+        async with database:
+            for search_query in search_queries:
+                results = await search_engine.search(
+                    query=search_query,
+                    limit=limit * 2,  # Get more results for LLM to analyze
+                    similarity_threshold=config.similarity_threshold,
+                    include_context=True,
+                )
+
+                # Post-filter results by file pattern if specified
+                if files and results:
+                    filtered_results = []
+                    for result in results:
+                        # Get relative path from project root
+                        try:
+                            rel_path = str(result.file_path.relative_to(project_root))
+                        except ValueError:
+                            # If file is outside project root, use absolute path
+                            rel_path = str(result.file_path)
+
+                        # Match against glob pattern (both full path and basename)
+                        if fnmatch(rel_path, files) or fnmatch(
+                            os.path.basename(rel_path), files
+                        ):
+                            filtered_results.append(result)
+                    results = filtered_results
+
+                search_results[search_query] = results
+                total_results += len(results)
+
+                console.print(
+                    f"  • [yellow]{search_query}[/yellow]: {len(results)} results"
+                )
+
+    except Exception as e:
+        logger.error(f"Search execution failed: {e}")
+        print_error(f"Search failed: {e}")
+        raise typer.Exit(1)
+
+    if total_results == 0:
+        print_warning("\n⚠️  No results found for any search query.")
+        print_info("\n[bold]Suggestions:[/bold]")
+        print_info("  • Try rephrasing your question")
+        print_info("  • Use more general terms")
+        print_info(
+            "  • Check if relevant files are indexed with [cyan]mcp-vector-search status[/cyan]"
+        )
+        raise typer.Exit(0)
+
+    # Step 3: Have LLM analyze and rank results
+    console.print(f"\n[cyan]🤖 Analyzing {total_results} results...[/cyan]")
+
+    try:
+        ranked_results = await llm_client.analyze_and_rank_results(
+            original_query=query,
+            search_results=search_results,
+            top_n=limit,
+        )
+
+        if not ranked_results:
+            print_warning("\n⚠️  LLM could not identify relevant results.")
+            raise typer.Exit(0)
+
+    except SearchError as e:
+        print_error(f"Result analysis failed: {e}")
+        # Fallback: show raw search results
+        print_warning("\nShowing raw search results instead...")
+        await _show_fallback_results(search_results, limit)
+        raise typer.Exit(1)
+
+    # Step 4: Display results with explanations
+    if json_output:
+        await _display_json_results(ranked_results)
+    else:
+        await _display_rich_results(ranked_results, query)
+
+
+async def _display_rich_results(
+    ranked_results: list[dict],
+    original_query: str,
+) -> None:
+    """Display results in rich formatted output.
+
+    Args:
+        ranked_results: List of ranked results with explanations
+        original_query: Original user query
+    """
+    from rich.panel import Panel
+    from rich.syntax import Syntax
+
+    console.print(
+        f"\n[bold cyan]🎯 Top Results for:[/bold cyan] [white]{original_query}[/white]\n"
+    )
+
+    for i, item in enumerate(ranked_results, 1):
+        result = item["result"]
+        relevance = item["relevance"]
+        explanation = item["explanation"]
+        query = item["query"]
+
+        # Determine relevance emoji and color
+        if relevance == "High":
+            relevance_emoji = "🟢"
+            relevance_color = "green"
+        elif relevance == "Medium":
+            relevance_emoji = "🟡"
+            relevance_color = "yellow"
+        else:
+            relevance_emoji = "🔴"
+            relevance_color = "red"
+
+        # Header with result number and file
+        console.print(f"[bold]📍 Result {i} of {len(ranked_results)}[/bold]")
+        console.print(
+            f"[cyan]📂 {result.file_path.relative_to(result.file_path.parent.parent)}[/cyan]"
+        )
+
+        # Relevance and explanation
+        console.print(
+            f"\n{relevance_emoji} [bold {relevance_color}]Relevance: {relevance}[/bold {relevance_color}]"
+        )
+        console.print(f"[dim]Search query: {query}[/dim]")
+        console.print(f"\n💡 [italic]{explanation}[/italic]\n")
+
+        # Code snippet with syntax highlighting
+        file_ext = result.file_path.suffix.lstrip(".")
+        code_syntax = Syntax(
+            result.content,
+            lexer=file_ext or "python",
+            theme="monokai",
+            line_numbers=True,
+            start_line=result.start_line,
+        )
+
+        panel = Panel(
+            code_syntax,
+            title=f"[bold]{result.function_name or result.class_name or 'Code'}[/bold]",
+            border_style="cyan",
+        )
+        console.print(panel)
+
+        # Metadata
+        metadata = []
+        if result.function_name:
+            metadata.append(f"Function: [cyan]{result.function_name}[/cyan]")
+        if result.class_name:
+            metadata.append(f"Class: [cyan]{result.class_name}[/cyan]")
+        metadata.append(f"Lines: [cyan]{result.start_line}-{result.end_line}[/cyan]")
+        metadata.append(f"Similarity: [cyan]{result.similarity_score:.3f}[/cyan]")
+
+        console.print("[dim]" + " | ".join(metadata) + "[/dim]")
+        console.print()  # Blank line between results
+
+    # Footer with tips
+    console.print("[dim]─" * 80 + "[/dim]")
+    console.print(
+        "\n[dim]💡 Tip: Try different phrasings or add more specific terms for better results[/dim]"
+    )
+
+
+async def _display_json_results(ranked_results: list[dict]) -> None:
+    """Display results in JSON format.
+
+    Args:
+        ranked_results: List of ranked results with explanations
+    """
+    from ..output import print_json
+
+    json_data = []
+    for item in ranked_results:
+        result = item["result"]
+        json_data.append(
+            {
+                "file": str(result.file_path),
+                "start_line": result.start_line,
+                "end_line": result.end_line,
+                "function_name": result.function_name,
+                "class_name": result.class_name,
+                "content": result.content,
+                "similarity_score": result.similarity_score,
+                "relevance": item["relevance"],
+                "explanation": item["explanation"],
+                "search_query": item["query"],
+            }
+        )
+
+    print_json(json_data, title="Chat Search Results")
+
+
+async def _show_fallback_results(
+    search_results: dict[str, list],
+    limit: int,
+) -> None:
+    """Show fallback results when LLM analysis fails.
+
+    Args:
+        search_results: Dictionary of search queries to results
+        limit: Number of results to show
+    """
+    from ..output import print_search_results
+
+    # Flatten and deduplicate results
+    all_results = []
+    seen_files = set()
+
+    for results in search_results.values():
+        for result in results:
+            file_key = (result.file_path, result.start_line)
+            if file_key not in seen_files:
+                all_results.append(result)
+                seen_files.add(file_key)
+
+    # Sort by similarity score
+    all_results.sort(key=lambda r: r.similarity_score, reverse=True)
+
+    # Show top N
+    print_search_results(
+        results=all_results[:limit],
+        query="Combined search results",
+        show_content=True,
+    )
+
+
+if __name__ == "__main__":
+    chat_app()