Model Context Protocol Demo

import streamlit as st import asyncio import json import logging import os import sqlite3 import pandas as pd import time from typing import Dict, List, Any, Optional, Tuple from fastmcp import Client from datetime import datetime import hashlib import numpy as np # RAG dependencies try: from sentence_transformers import SentenceTransformer from sentence_transformers import util as st_util from sklearn.metrics.pairwise import cosine_similarity RAG_AVAILABLE = True except ImportError: RAG_AVAILABLE = False st.warning("⚠️ RAG features disabled. Install: pip install sentence-transformers scikit-learn") # Configure logging logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s') # Streamlit page config st.set_page_config( page_title="MCP Client with RAG", page_icon="🧠", layout="wide", initial_sidebar_state="expanded" ) SAMPLE_QUERIES = """ - 15 + 27 - what is 3 times 4 - 9 divide 3 - calculate power of 2 to 3 - compute square root of 2 - sine of 30 degrees - tan(45 degree) - find arc tangent of 1 in degree - get stock price of GOOG - find company details for ticker GOOG - Get company info about AAPL - health check - server diagnostics - repeat this message: hello MCP server - what is pi (fail) """ CUSTOM_CSS_STYLE = """ <style> .main-header { font-size: 2.5rem; font-weight: bold; text-align: center; margin-bottom: 1rem; background: linear-gradient(90deg, #FF6B6B, #4ECDC4, #45B7D1); -webkit-background-clip: text; -webkit-text-fill-color: transparent; background-clip: text; } .tool-call { background-color: #e7f3ff; border-left: 4px solid #0066cc; padding: 0.75rem; margin: 0.5rem 0; border-radius: 0.375rem; } .error-message { background-color: #f8d7da; border-left: 4px solid #dc3545; padding: 0.75rem; margin: 0.5rem 0; border-radius: 0.375rem; } .debug-info { background-color: #fff3cd; border-left: 4px solid #ffc107; padding: 0.75rem; margin: 0.5rem 0; border-radius: 0.375rem; font-family: monospace; font-size: 0.9rem; } .rag-match { background-color: #d1ecf1; border-left: 4px solid #17a2b8; padding: 0.5rem; margin: 0.25rem 0; border-radius: 0.25rem; font-size: 0.9rem; } .similarity-score { background-color: #d4edda; color: #155724; padding: 0.2rem 0.4rem; border-radius: 0.2rem; font-weight: bold; font-size: 0.8rem; } </style> """ # Custom CSS st.markdown(CUSTOM_CSS_STYLE, unsafe_allow_html=True) # LLM Models and Provider Selection LLM_PROVIDER_MAP = { "google": [ "gemini-2.5-flash-preview-05-20", "gemini-2.5-pro-preview-05-06", "gemini-2.0-flash", "gemini-1.5-flash", "gemini-1.5-pro", ], "openai": [ "gpt-4o-mini", "gpt-4o", "gpt-3.5-turbo", ], "anthropic": [ "claude-3-5-sonnet-20241022", "claude-3-7-sonnet", ], # "ollama": [ # "qwen2.5", # ], # "bedrock": [ # "anthropic.claude-3-5-sonnet-20241022-v2:0", # ], } MCP_SERVER_PATH = "mcp_server.py" SQLITE_DB_FILE = "mcp_chat_history.db" TABLE_CHAT_HISTORY = "chat_history" CHAT_HISTORY_DDL = f""" CREATE TABLE IF NOT EXISTS {TABLE_CHAT_HISTORY} ( id INTEGER PRIMARY KEY AUTOINCREMENT, session_id TEXT NOT NULL, timestamp DATETIME NOT NULL, llm_provider TEXT, model_name TEXT, parsing_mode TEXT, user_query TEXT NOT NULL, parsed_action TEXT, tool_name TEXT, resource_uri TEXT, parameters TEXT, confidence REAL, reasoning TEXT, rag_matches TEXT, similarity_scores TEXT, response_data TEXT, formatted_response TEXT, elapsed_time_ms INTEGER, error_message TEXT, success BOOLEAN NOT NULL DEFAULT 1 ) """ def ask_llm(provider, client, model_name, query, system_prompt, max_tokens=300, temperature=0.1): # Get LLM response with dynamic prompt if provider == "anthropic": response = client.messages.create( model=model_name, max_tokens=max_tokens, temperature=temperature, system=system_prompt, messages=[{"role": "user", "content": f"Query: {query}"}] ) llm_response = response.content[0].text.strip() elif provider == "openai": response = client.chat.completions.create( model=model_name, messages=[ {"role": "system", "content": system_prompt}, {"role": "user", "content": f"Query: {query}"} ], temperature=temperature, max_tokens=max_tokens ) llm_response = response.choices[0].message.content.strip() elif provider == "google": response = client.generate_content( f"{system_prompt}\n\nUser Query: {query}", generation_config={ "temperature": temperature, "max_output_tokens": max_tokens, } ) llm_response = response.text.strip() else: st.error(f"Unsupported LLM provider: {provider}") return None # Clean and parse JSON if llm_response.startswith("```json"): llm_response = llm_response.replace("```json", "").replace("```", "").strip() elif llm_response.startswith("```"): llm_response = llm_response.replace("```", "").strip() parsed_response = json.loads(llm_response) return parsed_response # --- Database Operations --- class ChatHistoryDB: def __init__(self, db_file: str = SQLITE_DB_FILE): self.db_file = db_file self.init_database() def init_database(self): with sqlite3.connect(self.db_file) as conn: cursor = conn.cursor() # Create table with original schema first cursor.execute(CHAT_HISTORY_DDL) # # Check if new columns exist and add them if needed # cursor.execute("PRAGMA table_info(chat_history)") # columns = [column[1] for column in cursor.fetchall()] # if 'rag_matches' not in columns: # cursor.execute("ALTER TABLE chat_history ADD COLUMN rag_matches TEXT") # logging.info("✅ Added rag_matches column to database") # if 'similarity_scores' not in columns: # cursor.execute("ALTER TABLE chat_history ADD COLUMN similarity_scores TEXT") # logging.info("✅ Added similarity_scores column to database") cursor.execute(f"CREATE INDEX IF NOT EXISTS idx_session_id ON {TABLE_CHAT_HISTORY}(session_id)") conn.commit() def insert_chat_entry(self, entry: Dict[str, Any]) -> int: with sqlite3.connect(self.db_file) as conn: cursor = conn.cursor() cursor.execute(f""" INSERT INTO {TABLE_CHAT_HISTORY} ( session_id, timestamp, llm_provider, model_name, parsing_mode, user_query, parsed_action, tool_name, resource_uri, parameters, confidence, reasoning, rag_matches, similarity_scores, response_data, formatted_response, elapsed_time_ms, error_message, success ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?) """, ( entry.get('session_id'), entry.get('timestamp'), entry.get('llm_provider'), entry.get('model_name'), entry.get('parsing_mode'), entry.get('user_query'), entry.get('parsed_action'), entry.get('tool_name'), entry.get('resource_uri'), entry.get('parameters'), entry.get('confidence'), entry.get('reasoning'), entry.get('rag_matches'), entry.get('similarity_scores'), entry.get('response_data'), entry.get('formatted_response'), entry.get('elapsed_time_ms'), entry.get('error_message'), entry.get('success', True) )) entry_id = cursor.lastrowid conn.commit() return entry_id def get_chat_history(self, limit: int = 100, filters: Dict[str, Any] = None) -> List[Dict[str, Any]]: with sqlite3.connect(self.db_file) as conn: cursor = conn.cursor() query = f"SELECT * FROM {TABLE_CHAT_HISTORY}" params = [] if filters and filters.get('session_id'): query += " WHERE session_id = ?" params.append(filters['session_id']) query += " ORDER BY timestamp DESC LIMIT ?" params.append(limit) cursor.execute(query, params) columns = [description[0] for description in cursor.description] return [dict(zip(columns, row)) for row in cursor.fetchall()] # --- RAG System for MCP Tools/Resources --- class MCPRAGSystem: def __init__(self): self.model = None self.tool_embeddings = None self.resource_embeddings = None self.tool_contexts = [] self.resource_contexts = [] if RAG_AVAILABLE: self.initialize_model() def initialize_model(self): """Initialize sentence transformer model""" try: # Use a fast, efficient model for embeddings self.model = SentenceTransformer('all-MiniLM-L6-v2') logging.info("✅ RAG System initialized with all-MiniLM-L6-v2") except Exception as e: logging.error(f"❌ Failed to initialize RAG model: {e}") self.model = None def build_rich_context(self, item: Dict, item_type: str) -> str: """Build rich context for tools and resources with examples and synonyms""" if item_type == "tool": name = item.get('name', '') desc = item.get('description', '') # Enhanced context with usage examples and synonyms context_map = { 'calculator': """ Tool: calculator Description: Performs mathematical arithmetic operations Type: computation tool Usage examples: - Basic math: "15 plus 27", "multiply 8 by 4", "divide 100 by 5" - Advanced: "what's 2 to the power of 3", "square root calculation" - Keywords: add, subtract, multiply, divide, power, math, compute, calculate Synonyms: math, arithmetic, computation, calculate, compute """, 'trig': """ Tool: trig Description: Trigonometric functions (sine, cosine, tangent) Type: mathematical tool Usage examples: - "sine of 30 degrees", "cosine of 45", "tangent of 60 degrees" - "sin(π/4)", "cos(0)", "tan(90 degrees)" - Unit support: degrees, radians Keywords: trigonometry, sine, cosine, tangent, sin, cos, tan, angle Synonyms: trigonometry, trig functions, angles, geometry """, 'health': """ Tool: health Description: Server health check and status monitoring Type: diagnostic tool Usage examples: - "health check", "server status", "is server running" - "ping server", "system status", "connectivity test" Keywords: health, status, ping, check, monitor, diagnostic Synonyms: status, ping, check, monitor, diagnostic, connectivity """, 'echo': """ Tool: echo Description: Echo back messages for testing Type: utility tool Usage examples: - "echo hello world", "repeat this message", "say hello" - Testing connectivity and response Keywords: echo, repeat, say, message, test Synonyms: repeat, say, message, test, respond """ } # Return specific context or generic one return context_map.get(name, f""" Tool: {name} Description: {desc} Type: generic tool Usage: General purpose tool for {name} operations Keywords: {name} """).strip() elif item_type == "resource": # Handle both string URIs and AnyUrl objects safely uri_raw = item.get('uri', '') try: # Convert AnyUrl to string if needed uri = str(uri_raw) if hasattr(uri_raw, '__str__') else uri_raw except Exception: uri = 'unknown_resource' desc = item.get('description', '') # Build context for resources - use safe string operations try: uri_lower = uri.lower() if isinstance(uri, str) else str(uri).lower() if 'stock' in uri_lower: return f""" Resource: {uri} Description: {desc} Type: financial data resource Usage examples: - Stock information, financial data, company details - Market data, stock prices, financial analysis Keywords: stock, finance, market, company, financial, investment Synonyms: stocks, shares, equity, financial data, market data """.strip() else: # Safe extraction of resource name try: resource_name = uri.split('/')[-1] if '/' in str(uri) else str(uri) except Exception: resource_name = 'resource' return f""" Resource: {uri} Description: {desc} Type: data resource Usage: Access to {uri} data and information Keywords: {resource_name} """.strip() except Exception as e: # Fallback for any string operation failures logging.warning(f"Failed to process resource URI: {e}") return f""" Resource: {uri} Description: {desc} Type: data resource Usage: General data resource Keywords: data, resource """.strip() return f"{item_type}: {item}" def build_embeddings(self, tools: List[Dict], resources: List[Dict]): """Build embeddings for all tools and resources""" if not self.model: return try: # Build rich contexts with error handling self.tool_contexts = [] for tool in tools: try: context = self.build_rich_context(tool, 'tool') self.tool_contexts.append({ 'name': tool.get('name', ''), 'description': tool.get('description', ''), 'context': context, 'type': 'tool' }) except Exception as e: logging.warning(f"Failed to build context for tool {tool.get('name', 'unknown')}: {e}") # Add fallback context self.tool_contexts.append({ 'name': tool.get('name', 'unknown'), 'description': tool.get('description', ''), 'context': f"Tool: {tool.get('name', 'unknown')}\nDescription: {tool.get('description', '')}", 'type': 'tool' }) self.resource_contexts = [] for resource in resources: try: context = self.build_rich_context(resource, 'resource') # Safe URI extraction uri_raw = resource.get('uri', '') uri = str(uri_raw) if uri_raw else 'unknown_resource' self.resource_contexts.append({ 'uri': uri, 'description': resource.get('description', ''), 'context': context, 'type': 'resource' }) except Exception as e: logging.warning(f"Failed to build context for resource {resource.get('uri', 'unknown')}: {e}") # Add fallback context uri_raw = resource.get('uri', 'unknown_resource') uri = str(uri_raw) if uri_raw else 'unknown_resource' self.resource_contexts.append({ 'uri': uri, 'description': resource.get('description', ''), 'context': f"Resource: {uri}\nDescription: {resource.get('description', '')}", 'type': 'resource' }) # Create embeddings with error handling if self.tool_contexts: try: tool_texts = [item['context'] for item in self.tool_contexts] self.tool_embeddings = self.model.encode(tool_texts) logging.info(f"✅ Built embeddings for {len(self.tool_contexts)} tools") except Exception as e: logging.error(f"❌ Failed to encode tool embeddings: {e}") self.tool_embeddings = None if self.resource_contexts: try: resource_texts = [item['context'] for item in self.resource_contexts] self.resource_embeddings = self.model.encode(resource_texts) logging.info(f"✅ Built embeddings for {len(self.resource_contexts)} resources") except Exception as e: logging.error(f"❌ Failed to encode resource embeddings: {e}") self.resource_embeddings = None except Exception as e: logging.error(f"❌ Failed to build embeddings: {e}") # Ensure we have empty but valid states self.tool_contexts = [] self.resource_contexts = [] self.tool_embeddings = None self.resource_embeddings = None def semantic_search(self, query: str, top_k: int = 5) -> List[Dict]: """Perform optimized semantic search across tools and resources using sentence-transformers util""" if not self.model: return [] try: # Encode the query query_embedding = self.model.encode([query]) # Combine tool and resource embeddings efficiently all_embeddings = [] all_contexts = [] if self.tool_embeddings is not None and len(self.tool_embeddings) > 0: all_embeddings.append(self.tool_embeddings) all_contexts.extend([(ctx, 'tool') for ctx in self.tool_contexts]) if self.resource_embeddings is not None and len(self.resource_embeddings) > 0: all_embeddings.append(self.resource_embeddings) all_contexts.extend([(ctx, 'resource') for ctx in self.resource_contexts]) if not all_embeddings: return [] # Concatenate all embeddings into single corpus corpus_embeddings = np.concatenate(all_embeddings, axis=0) # Use optimized semantic_search from sentence-transformers search_results = st_util.semantic_search( query_embedding, corpus_embeddings, top_k=top_k, score_function=st_util.cos_sim )[0] # Get results for first (and only) query # Map results back to our context format results = [] for hit in search_results: corpus_id = hit['corpus_id'] similarity = float(hit['score']) # Only include results above minimum threshold if similarity > 0.1: context_item, item_type = all_contexts[corpus_id] results.append({ 'item': context_item, 'similarity': similarity, 'type': item_type }) logging.info(f"✅ Semantic search found {len(results)} relevant items for query: '{query[:50]}...'") return results except Exception as e: logging.error(f"❌ Optimized semantic search failed: {e}") # Fallback to original method if needed return self._fallback_semantic_search(query, top_k) def _fallback_semantic_search(self, query: str, top_k: int = 5) -> List[Dict]: """Fallback to original semantic search method if optimized version fails""" if not self.model: return [] try: # Encode the query query_embedding = self.model.encode([query]) results = [] # Search tools using original method if self.tool_embeddings is not None and len(self.tool_embeddings) > 0: tool_similarities = cosine_similarity(query_embedding, self.tool_embeddings)[0] for i, similarity in enumerate(tool_similarities): if similarity > 0.1: # Minimum similarity threshold results.append({ 'item': self.tool_contexts[i], 'similarity': float(similarity), 'type': 'tool' }) # Search resources using original method if self.resource_embeddings is not None and len(self.resource_embeddings) > 0: resource_similarities = cosine_similarity(query_embedding, self.resource_embeddings)[0] for i, similarity in enumerate(resource_similarities): if similarity > 0.1: # Minimum similarity threshold results.append({ 'item': self.resource_contexts[i], 'similarity': float(similarity), 'type': 'resource' }) # Sort by similarity and return top_k results.sort(key=lambda x: x['similarity'], reverse=True) logging.warning(f"⚠️ Used fallback search method for query: '{query[:50]}...'") return results[:top_k] except Exception as e: logging.error(f"❌ Fallback semantic search also failed: {e}") return [] def build_dynamic_prompt(self, relevant_items: List[Dict], query: str) -> str: """Build dynamic system prompt based on relevant items""" if not relevant_items: return """ You are a tool selection assistant. Respond with ONLY a JSON object with action, tool, params, confidence, and reasoning fields. """ # Build tools section tools_section = "Available tools:\n" for item in relevant_items: if item['type'] == 'tool': tool_info = item['item'] similarity = item['similarity'] tools_section += f"- {tool_info['name']}: {tool_info['description']} (relevance: {similarity:.2f})\n" # Build resources section resources_section = "\nAvailable resources:\n" for item in relevant_items: if item['type'] == 'resource': resource_info = item['item'] similarity = item['similarity'] resources_section += f"- {resource_info['uri']}: {resource_info['description']} (relevance: {similarity:.2f})\n" # Examples based on most relevant items examples_section = "\nExamples based on available tools:\n" # Add specific examples for discovered tools for item in relevant_items[:3]: # Top 3 most relevant if item['type'] == 'tool': tool_name = item['item']['name'] if tool_name == 'calculator': examples_section += '- "15 plus 27" -> {"action": "tool", "tool": "calculator", "params": {"operation": "add", "num1": 15, "num2": 27}, "confidence": 0.98, "reasoning": "Simple addition"}\n' elif tool_name == 'trig': examples_section += '- "sine of 30 degrees" -> {"action": "tool", "tool": "trig", "params": {"operation": "sine", "num1": 30, "unit": "degree"}, "confidence": 0.95, "reasoning": "Trigonometric calculation"}\n' elif tool_name == 'health': examples_section += '- "health check" -> {"action": "tool", "tool": "health", "params": {}, "confidence": 0.9, "reasoning": "Server health check"}\n' elif tool_name == 'echo': examples_section += '- "echo hello" -> {"action": "tool", "tool": "echo", "params": {"message": "hello"}, "confidence": 0.95, "reasoning": "Echo command"}\n' system_prompt = f""" You are an intelligent tool selection assistant. Analyze the user query and respond with ONLY a JSON object: {{ "action": "tool", "tool": "tool_name_or_null", "params": {{"param1": "value1"}}, "confidence": 0.95, "reasoning": "Brief explanation" }} {tools_section} {resources_section} {examples_section} Instructions: - Only use tools/resources listed above - Consider the relevance scores when making decisions - Set confidence based on query clarity and tool match - If no tool matches well (all relevance < 0.3), set tool to null - Respond with ONLY the JSON object, no other text. """ return system_prompt # Initialize session state def init_session_state(): if 'chat_history_db' not in st.session_state: st.session_state.chat_history_db = ChatHistoryDB() if 'session_id' not in st.session_state: st.session_state.session_id = hashlib.md5(f"{datetime.now()}{os.getpid()}".encode()).hexdigest()[:8] if 'llm_provider' not in st.session_state: st.session_state.llm_provider = "google" if 'use_llm' not in st.session_state: st.session_state.use_llm = True if 'server_connected' not in st.session_state: st.session_state.server_connected = False if 'available_tools' not in st.session_state: st.session_state.available_tools = [] if 'available_resources' not in st.session_state: st.session_state.available_resources = [] if 'use_rag' not in st.session_state: st.session_state.use_rag = True if 'last_parsed_query' not in st.session_state: st.session_state.last_parsed_query = None if 'rag_system' not in st.session_state: st.session_state.rag_system = MCPRAGSystem() if 'last_rag_matches' not in st.session_state: st.session_state.last_rag_matches = [] # --- Enhanced LLM Query Parser with RAG --- class LLMQueryParser: def __init__(self, provider: str = "google"): self.provider = provider self.client = None self.model_name = None self.setup_llm_client() def setup_llm_client(self): try: if self.provider == "anthropic": import anthropic api_key = os.getenv("ANTHROPIC_API_KEY") if api_key: self.client = anthropic.Anthropic(api_key=api_key) self.model_name = st.session_state.llm_model_name elif self.provider == "openai": import openai api_key = os.getenv("OPENAI_API_KEY") if api_key: self.client = openai.OpenAI(api_key=api_key) self.model_name = st.session_state.llm_model_name elif self.provider == "google": import google.generativeai as genai api_key = os.getenv("GEMINI_API_KEY") if api_key: genai.configure(api_key=api_key) self.model_name = st.session_state.llm_model_name self.client = genai.GenerativeModel(self.model_name) except Exception as e: st.error(f"Failed to initialize {self.provider}: {e}") self.client = None def parse_query_with_rag(self, query: str, rag_system: MCPRAGSystem) -> Optional[Dict[str, Any]]: """Parse query using RAG-enhanced semantic search""" if not self.client or not rag_system.model: return None try: # Perform semantic search relevant_items = rag_system.semantic_search(query, top_k=5) # Store RAG matches for debugging st.session_state.last_rag_matches = relevant_items if not relevant_items: # Fallback to standard parsing if no relevant items found return self.parse_query_sync(query) # Build dynamic prompt based on relevant items system_prompt = rag_system.build_dynamic_prompt(relevant_items, query) parsed_response = ask_llm(self.provider, self.client, self.model_name, query, system_prompt) if not parsed_response: return None # Add RAG metadata parsed_response['rag_enhanced'] = True parsed_response['rag_matches'] = len(relevant_items) if parsed_response.get("action") and parsed_response.get("confidence", 0) >= 0.3: return parsed_response except Exception as e: logging.error(f"RAG-enhanced parsing error: {e}") # Fallback to standard parsing return self.parse_query_sync(query) return None def parse_query_sync(self, query: str) -> Optional[Dict[str, Any]]: """Legacy parsing method with hardcoded examples""" if not self.client: return None system_prompt = """ You are a tool selection assistant. Respond with ONLY a JSON object: { "action": "tool", "tool": "tool_name_or_null", "params": {"param1": "value1"}, "confidence": 0.95, "reasoning": "Brief explanation" } Available tools: - calculator: operation (add/subtract/multiply/divide/power), num1, num2 - trig: operation (sine/cosine/tangent), num1, unit (degree/radian) - health: no parameters - echo: message Examples: "15 plus 27" -> {"action": "tool", "tool": "calculator", "params": {"operation": "add", "num1": 15, "num2": 27}, "confidence": 0.98, "reasoning": "Simple addition"} "sine of 30 degrees" -> {"action": "tool", "tool": "trig", "params": {"operation": "sine", "num1": 30, "unit": "degree"}, "confidence": 0.95, "reasoning": "Trigonometric calculation"} Respond with ONLY the JSON object. """ try: parsed_response = ask_llm(self.provider, self.client, self.model_name, query, system_prompt) if not parsed_response: return None parsed_response['rag_enhanced'] = False if parsed_response.get("action") and parsed_response.get("confidence", 0) >= 0.5: return parsed_response except Exception as e: st.error(f"LLM parsing error: {e}") return None # --- Rule-based Parser --- class RuleBasedQueryParser: @staticmethod def parse_query(query: str) -> Optional[Dict[str, Any]]: import re query_lower = query.lower().strip() # Health check if any(word in query_lower for word in ["health", "status", "ping"]): return {"action": "tool", "tool": "health", "params": {}, "confidence": 0.9, "reasoning": "Health check request", "rag_enhanced": False} # Echo command if query_lower.startswith("echo "): return {"action": "tool", "tool": "echo", "params": {"message": query[5:].strip()}, "confidence": 0.95, "reasoning": "Echo command", "rag_enhanced": False} # Calculator calc_patterns = [ ("add", ["plus", "add", "+", "sum"]), ("subtract", ["minus", "subtract", "-"]), ("multiply", ["times", "multiply", "*", "×"]), ("divide", ["divide", "divided by", "/"]), ("power", ["power", "to the power", "^"]) ] for operation, keywords in calc_patterns: for keyword in keywords: if keyword in query_lower: numbers = re.findall(r'-?\d+(?:\.\d+)?', query) if len(numbers) >= 2: return { "action": "tool", "tool": "calculator", "params": {"operation": operation, "num1": float(numbers[0]), "num2": float(numbers[1])}, "confidence": 0.9, "reasoning": f"Calculator operation: {operation}", "rag_enhanced": False } # Trig functions trig_patterns = [ ("sine", ["sine", "sin"]), ("cosine", ["cosine", "cos"]), ("tangent", ["tangent", "tan"]) ] for operation, keywords in trig_patterns: for keyword in keywords: if keyword in query_lower: numbers = re.findall(r'-?\d+(?:\.\d+)?', query) if numbers: unit = "radian" if any(word in query_lower for word in ["radian", "rad"]) else "degree" return { "action": "tool", "tool": "trig", "params": {"operation": operation, "num1": float(numbers[0]), "unit": unit}, "confidence": 0.9, "reasoning": f"Trigonometry: {operation}", "rag_enhanced": False } return None # --- Utility Functions --- def extract_result_data(result): try: if isinstance(result, list) and len(result) > 0: content_item = result[0] if hasattr(content_item, 'text'): try: return json.loads(content_item.text) except json.JSONDecodeError: return {"text": content_item.text} else: return {"content": str(content_item)} elif hasattr(result, 'content') and result.content: content_item = result.content[0] if hasattr(content_item, 'text'): try: return json.loads(content_item.text) except json.JSONDecodeError: return {"text": content_item.text} else: return {"content": str(content_item)} else: return result if isinstance(result, dict) else {"result": str(result)} except Exception as e: return {"error": f"Could not parse result: {e}"} def format_result_for_display(tool_name: str, result: Dict) -> str: if isinstance(result, dict) and "error" in result: return f"❌ [Error] {result['error']}" if tool_name == "calculator": expression = result.get('expression', f"{result.get('num1', '?')} {result.get('operation', '?')} {result.get('num2', '?')} = {result.get('result', '?')}") return f"🧮 [Calculator] {expression}" elif tool_name == "trig": expression = result.get('expression', f"{result.get('operation', '?')}({result.get('num1', '?')}) = {result.get('result', '?')}") return f"📐 [Trigonometry] {expression}" elif tool_name == "health": return f"✅ [Health] {result.get('message', 'Server is healthy')}" elif tool_name == "echo": return f"🔊 [Echo] {result.get('echo', result.get('message', str(result)))}" return f"✅ [Result] {json.dumps(result, indent=2)}" # --- CACHED MCP Operations using st.cache_resource --- @st.cache_resource def get_mcp_server_info(): """Get cached server info (tools/resources) - cached across reruns""" async def _discover(): async with Client(MCP_SERVER_PATH) as client: # Get tools tools = await client.list_tools() available_tools = [{"name": tool.name, "description": tool.description} for tool in tools] if tools else [] # Get resources try: resources = await client.list_resources() available_resources = [{"uri": resource.uri, "description": resource.description} for resource in resources] if resources else [] except: available_resources = [] return available_tools, available_resources return asyncio.run(_discover()) async def execute_mcp_query_async(parsed_query): """Execute MCP query with proper async context manager""" start_time = time.time() action = parsed_query.get("action") tool_name = parsed_query.get("tool") parameters = parsed_query.get("params", {}) results = [] if action == "tool" and tool_name: try: # Use proper async context manager for each query async with Client(MCP_SERVER_PATH) as client: tool_result = await client.call_tool(tool_name, parameters) tool_data = extract_result_data(tool_result) results.append({ "type": "tool", "name": tool_name, "data": tool_data, "success": "error" not in tool_data }) except Exception as e: results.append({ "type": "error", "message": f"Tool call error: {e}", "success": False }) elapsed_time = int((time.time() - start_time) * 1000) return results, elapsed_time def do_sidebar(): with st.sidebar: st.header("⚙️ Configuration") st.info(f"📍 [Session ID] `{st.session_state.session_id}`") # RAG System Status st.subheader("🧠 RAG System") if RAG_AVAILABLE and st.session_state.rag_system.model: st.success("✅ RAG System Active") st.info("🔍 Semantic search enabled") # RAG settings st.session_state.use_rag = st.checkbox( "🎯 Use RAG-Enhanced Parsing", value=st.session_state.use_rag, help="Use semantic search to find relevant tools dynamically" ) else: st.error("❌ RAG System Disabled") st.warning("Install: `pip install sentence-transformers scikit-learn`") st.session_state.use_rag = False # LLM Provider/Model Selection c1, c2 = st.columns([2,3]) with c1: LLM_PROVIDER_LIST = list(LLM_PROVIDER_MAP.keys()) st.session_state.llm_provider = st.selectbox( "🤖 LLM Provider", LLM_PROVIDER_LIST, index=LLM_PROVIDER_LIST.index("google") ) with c2: LLM_MODEL_NAME_LIST = LLM_PROVIDER_MAP.get(st.session_state.llm_provider) st.session_state.llm_model_name = st.selectbox( "Model Name", LLM_MODEL_NAME_LIST, index=0 ) # Parsing Mode st.session_state.use_llm = st.checkbox( "🧠 Use LLM Parsing", value=st.session_state.use_llm ) # Parsing Mode Display if st.session_state.use_llm and st.session_state.use_rag and RAG_AVAILABLE: st.info("🎯 [Mode] RAG-enhanced LLM") elif st.session_state.use_llm: st.info("🤖 [Mode] LLM-based") else: st.info("📝 [Mode] Rule-based") # API Keys Status st.subheader("🔑 API Keys Status") api_keys_status = { "OpenAI": "✅" if os.getenv("OPENAI_API_KEY") else "❌", "Anthropic": "✅" if os.getenv("ANTHROPIC_API_KEY") else "❌", "Google": "✅" if os.getenv("GEMINI_API_KEY") else "❌", } for provider, status in api_keys_status.items(): st.write(f"{status} {provider}") # Server Connection using st.cache_resource for discovery only! st.subheader("🔌 Server Status") # Try to get cached server info (tools/resources) try: # This will use cached discovery if available tools, resources = get_mcp_server_info() # If we get here, server is reachable st.session_state.server_connected = True st.session_state.available_tools = tools st.session_state.available_resources = resources # Build RAG embeddings when tools/resources are available if RAG_AVAILABLE and st.session_state.rag_system.model: st.session_state.rag_system.build_embeddings(tools, resources) except Exception as e: st.session_state.server_connected = False st.session_state.available_tools = [] st.session_state.available_resources = [] if st.button("🔄 Refresh Server Discovery"): # Clear cache and rediscover st.cache_resource.clear() st.rerun() # Connection Status if st.session_state.server_connected: st.success("🟢 Server Connected") # Show tools and resources if connected if st.session_state.available_tools: with st.expander("🔧 Available Tools"): for tool in st.session_state.available_tools: st.write(f"• [{tool['name']}] {tool['description']}") if st.session_state.available_resources: with st.expander("📚 Available Resources"): for resource in st.session_state.available_resources: st.write(f"• [{resource['uri']}] {resource['description']}") # RAG Embeddings Status if RAG_AVAILABLE and st.session_state.rag_system.model: tool_count = len(st.session_state.rag_system.tool_contexts) resource_count = len(st.session_state.rag_system.resource_contexts) st.info(f"🎯 RAG: {tool_count} tools, {resource_count} resources indexed. \nDynamic resource is unavailable via list_resources()") else: st.error("🔴 Server Disconnected") st.info(f"💡 Make sure {MCP_SERVER_PATH} is running, then click 'Refresh Server Discovery'") # Example queries st.subheader("💡 Example Queries") st.markdown(SAMPLE_QUERIES) # if st.button("15 + 27"): # st.session_state.example_query = "15 + 27" # --- Main App --- def main(): init_session_state() # Header st.markdown('<h1 class="main-header">🧠 MCP Client with RAG</h1>', unsafe_allow_html=True) # Sidebar Configuration do_sidebar() # Main Content col1, col2 = st.columns([2, 1]) with col1: st.subheader("💬 Query Interface") # Query Input default_query = st.session_state.get('example_query', '') user_query = st.text_input( "🎯 Enter your query:", value=default_query, placeholder="compute the square root of 144" ) # Clear example after using it if 'example_query' in st.session_state: del st.session_state.example_query col_submit, col_clear = st.columns([1, 1]) with col_submit: submit_button = st.button("🚀 Submit Query", type="primary") with col_clear: if st.button("🗑️ Clear Session"): st.session_state.session_id = hashlib.md5(f"{datetime.now()}{os.getpid()}".encode()).hexdigest()[:8] st.session_state.last_parsed_query = None st.session_state.last_rag_matches = [] st.success("✅ New session started!") st.rerun() # Process Query using RAG! if submit_button and user_query: try: # Parse query parsed_query = None model_name = None rag_matches = [] if st.session_state.use_llm: parser = LLMQueryParser(st.session_state.llm_provider) if parser.client: # Use RAG-enhanced parsing if available if st.session_state.use_rag and RAG_AVAILABLE and st.session_state.rag_system.model: parsed_query = parser.parse_query_with_rag(user_query, st.session_state.rag_system) rag_matches = st.session_state.last_rag_matches else: parsed_query = parser.parse_query_sync(user_query) model_name = parser.model_name else: st.warning("🔄 LLM not available, using rule-based parsing") parsed_query = RuleBasedQueryParser.parse_query(user_query) else: parsed_query = RuleBasedQueryParser.parse_query(user_query) if parsed_query: # Store for debug display st.session_state.last_parsed_query = parsed_query # Execute query with proper async context manager results, elapsed_time = asyncio.run(execute_mcp_query_async(parsed_query)) st.session_state.elapsed_time = elapsed_time # Auto-update connection status if successful if results and any(r.get('success', True) for r in results): st.session_state.server_connected = True # Save to database with RAG data db_entry = { 'session_id': st.session_state.session_id, 'timestamp': datetime.now(), 'llm_provider': st.session_state.llm_provider if st.session_state.use_llm else None, 'model_name': model_name, 'parsing_mode': 'RAG-Enhanced' if parsed_query.get('rag_enhanced') else ('LLM' if st.session_state.use_llm else 'Rule-based'), 'user_query': user_query, 'parsed_action': parsed_query.get('action'), 'tool_name': parsed_query.get('tool'), 'parameters': json.dumps(parsed_query.get('params', {})), 'confidence': parsed_query.get('confidence'), 'reasoning': parsed_query.get('reasoning'), 'rag_matches': json.dumps([{ 'name': m['item'].get('name', m['item'].get('uri', '')), 'similarity': m['similarity'], 'type': m['type'] } for m in rag_matches]) if rag_matches else None, 'similarity_scores': json.dumps([m['similarity'] for m in rag_matches]) if rag_matches else None, 'elapsed_time_ms': elapsed_time, 'success': all(result.get('success', True) for result in results) } entry_id = st.session_state.chat_history_db.insert_chat_entry(db_entry) st.session_state.entry_id = entry_id # Show RAG matches if available if rag_matches: st.markdown("### 🎯 RAG Search Results:") for i, match in enumerate(rag_matches[:3]): # Show top 3 item = match['item'] similarity = match['similarity'] item_type = match['type'] if item_type == 'tool': name = item.get('name', 'Unknown') desc = item.get('description', '') else: name = item.get('uri', 'Unknown') desc = item.get('description', '') st.markdown(f""" <div class="rag-match"> <strong>#{i+1} {item_type.title()}:</strong> {name}<br> <small>{desc}</small><br> <span class="similarity-score">Similarity: {similarity:.3f}</span> </div> """, unsafe_allow_html=True) for result in results: if result['type'] == 'tool': formatted_display = format_result_for_display(result['name'], result['data']) st.markdown(f'<div class="tool-call">{formatted_display}</div>', unsafe_allow_html=True) elif result['type'] == 'error': st.markdown(f'<div class="error-message">❌ {result["message"]}</div>', unsafe_allow_html=True) else: st.error("❓ Failed to call MCP tool because I couldn't understand your query.") except Exception as e: st.error(f"❌ Error processing query: {e}") st.info("💡 Try clicking 'Refresh Server Discovery' if connection issues persist") with col2: st.subheader("📊 Query Analysis") # Display debug info if st.session_state.last_parsed_query: parsed_query = st.session_state.last_parsed_query parsing_mode = "RAG-Enhanced" if parsed_query.get('rag_enhanced') else "Legacy" st.success(f"✅ Query processed in {st.session_state.elapsed_time}ms using {parsing_mode} parsing (Entry ID: {st.session_state.entry_id})") st.markdown('<div class="debug-info">', unsafe_allow_html=True) st.markdown("🔍 Debug - Parsed Query:") debug_info = { "Action": parsed_query.get('action'), "Tool": parsed_query.get('tool'), "Parameters": parsed_query.get('params', {}), "Confidence": parsed_query.get('confidence'), "Reasoning": parsed_query.get('reasoning'), "RAG Enhanced": parsed_query.get('rag_enhanced', False), "RAG Matches": parsed_query.get('rag_matches', 0) } st.json(debug_info) st.markdown('</div>', unsafe_allow_html=True) # RAG matches details if st.session_state.last_rag_matches: with st.expander("🎯 Detailed RAG Matches"): for i, match in enumerate(st.session_state.last_rag_matches): item = match['item'] similarity = match['similarity'] item_type = match['type'] st.write(f"[Match #{i+1} ({item_type})]") st.write(f"• Similarity: {similarity:.4f}") if item_type == 'tool': st.write(f"• Tool: {item.get('name', 'Unknown')}") st.write(f"• Description: {item.get('description', 'No description')}") else: st.write(f"• Resource: {item.get('uri', 'Unknown')}") st.write(f"• Description: {item.get('description', 'No description')}") st.write("---") # Session stats try: recent_entries = st.session_state.chat_history_db.get_chat_history( limit=5, filters={'session_id': st.session_state.session_id} ) if recent_entries: with st.expander("[Session Statistics]"): latest_entry = recent_entries[0] st.info(f"🔍 [Parser] {latest_entry['parsing_mode']}") if latest_entry['model_name']: st.info(f"🤖 [Model] {latest_entry['model_name']}") if len(recent_entries) > 1: successful = sum(1 for entry in recent_entries if entry['success']) avg_time = sum(entry['elapsed_time_ms'] or 0 for entry in recent_entries) / len(recent_entries) rag_enhanced_count = sum(1 for entry in recent_entries if entry['parsing_mode'] == 'RAG-Enhanced') st.metric("Queries", len(recent_entries)) st.metric("Success Rate", f"{(successful/len(recent_entries)*100):.1f}%") st.metric("Avg Response Time", f"{avg_time:.0f}ms") st.metric("RAG-Enhanced", f"{rag_enhanced_count}/{len(recent_entries)}") else: st.info("💡 No queries in this session yet. Try asking something!") except Exception as e: st.error(f"Error loading query analysis: {e}") if __name__ == "__main__": main()