import os, re, json, ast, asyncio
import pandas as pd
import streamlit as st
import base64
from io import BytesIO
from PIL import Image
from langchain_groq import ChatGroq
from langchain_core.messages import HumanMessage, SystemMessage
from fastmcp import Client
from fastmcp.client.transports import StreamableHttpTransport
import streamlit.components.v1 as components
import re
from dotenv import load_dotenv
load_dotenv()
# Initialize Groq client with environment variable
GROQ_API_KEY = os.environ.get("GROQ_API_KEY")
if not GROQ_API_KEY:
st.error("🔐 GROQ_API_KEY environment variable is not set. Please add it to your environment.")
st.stop()
groq_client = ChatGroq(
groq_api_key=GROQ_API_KEY,
model_name=os.environ.get("GROQ_MODEL", "openai/gpt-oss-20b")
)
# ========== PAGE CONFIG ==========
st.set_page_config(page_title="MCP CRUD Chat", layout="wide")
# ========== GLOBAL CSS ==========
st.markdown("""
<style>
[data-testid="stSidebar"] {
background: linear-gradient(180deg, #4286f4 0%, #397dd2 100%);
color: #fff !important;
min-width: 330px !important;
padding: 0 0 0 0 !important;
}
[data-testid="stSidebar"] .sidebar-title {
color: #fff !important;
font-weight: bold;
font-size: 2.2rem;
letter-spacing: -1px;
text-align: center;
margin-top: 28px;
margin-bottom: 18px;
}
.sidebar-block {
width: 94%;
margin: 0 auto 18px auto;
}
.sidebar-block label {
color: #fff !important;
font-weight: 500;
font-size: 1.07rem;
margin-bottom: 4px;
margin-left: 2px;
display: block;
text-align: left;
}
.sidebar-block .stSelectbox>div {
background: #fff !important;
color: #222 !important;
border-radius: 13px !important;
font-size: 1.13rem !important;
min-height: 49px !important;
box-shadow: 0 3px 14px #0002 !important;
padding: 3px 10px !important;
margin-top: 4px !important;
margin-bottom: 0 !important;
}
.stButton>button {
width: 100%;
height: 3rem;
background: #39e639;
color: #222;
font-size: 1.1rem;
font-weight: bold;
border-radius: 10px;
margin-bottom: 2rem;
}
/* Small refresh button styling */
.small-refresh-button button {
width: auto !important;
height: 2rem !important;
background: #4286f4 !important;
color: #fff !important;
font-size: 0.85rem !important;
font-weight: 500 !important;
border-radius: 6px !important;
margin-bottom: 0.5rem !important;
padding: 0.25rem 0.75rem !important;
border: none !important;
box-shadow: 0 2px 4px rgba(0,0,0,0.1) !important;
}
.small-refresh-button button:hover {
background: #397dd2 !important;
transform: translateY(-1px) !important;
box-shadow: 0 4px 8px rgba(0,0,0,0.15) !important;
}
.sidebar-logo-label {
margin-top: 30px !important;
margin-bottom: 10px;
font-size: 1.13rem !important;
font-weight: 600;
text-align: center;
color: #fff !important;
letter-spacing: 0.1px;
}
.sidebar-logo-row {
display: flex;
flex-direction: row;
justify-content: center;
align-items: center;
gap: 20px;
margin-top: 8px;
margin-bottom: 8px;
}
.sidebar-logo-row img {
width: 42px;
height: 42px;
border-radius: 9px;
background: #fff;
padding: 6px 8px;
object-fit: contain;
box-shadow: 0 2px 8px #0002;
}
/* Chat area needs bottom padding so sticky bar does not overlap */
.stChatPaddingBottom { padding-bottom: 98px; }
/* Responsive sticky chatbar */
.sticky-chatbar {
position: fixed;
left: 330px;
right: 0;
bottom: 0;
z-index: 100;
background: #f8fafc;
padding: 0.6rem 2rem 0.8rem 2rem;
box-shadow: 0 -2px 24px #0001;
}
@media (max-width: 800px) {
.sticky-chatbar { left: 0; right: 0; padding: 0.6rem 0.5rem 0.8rem 0.5rem; }
[data-testid="stSidebar"] { display: none !important; }
}
.chat-bubble {
padding: 13px 20px;
margin: 8px 0;
border-radius: 18px;
max-width: 75%;
font-size: 1.09rem;
line-height: 1.45;
box-shadow: 0 1px 4px #0001;
display: inline-block;
word-break: break-word;
}
.user-msg {
background: #e6f0ff;
color: #222;
margin-left: 24%;
text-align: right;
border-bottom-right-radius: 6px;
border-top-right-radius: 24px;
}
.agent-msg {
background: #f5f5f5;
color: #222;
margin-right: 24%;
text-align: left;
border-bottom-left-radius: 6px;
border-top-left-radius: 24px;
}
.chat-row {
display: flex;
align-items: flex-end;
margin-bottom: 0.6rem;
}
.avatar {
height: 36px;
width: 36px;
border-radius: 50%;
margin: 0 8px;
object-fit: cover;
box-shadow: 0 1px 4px #0002;
}
.user-avatar { order: 2; }
.agent-avatar { order: 0; }
.user-bubble { order: 1; }
.agent-bubble { order: 1; }
.right { justify-content: flex-end; }
.left { justify-content: flex-start; }
.chatbar-claude {
display: flex;
align-items: center;
gap: 12px;
width: 100%;
max-width: 850px;
margin: 0 auto;
border-radius: 20px;
background: #fff;
box-shadow: 0 2px 8px #0002;
padding: 8px 14px;
margin-bottom: 0;
}
.claude-hamburger {
background: #f2f4f9;
border: none;
border-radius: 11px;
font-size: 1.35rem;
font-weight: bold;
width: 38px; height: 38px;
cursor: pointer;
display: flex; align-items: center; justify-content: center;
transition: background 0.13s;
}
.claude-hamburger:hover { background: #e6f0ff; }
.claude-input {
flex: 1;
border: none;
outline: none;
font-size: 1.12rem;
padding: 0.45rem 0.5rem;
background: #f5f7fa;
border-radius: 8px;
min-width: 60px;
}
.claude-send {
background: #fe3044 !important;
color: #fff !important;
border: none;
border-radius: 50%;
width: 40px; height: 40px;
font-size: 1.4rem !important;
cursor: pointer;
display: flex; align-items: center; justify-content: center;
transition: background 0.17s;
}
.claude-send:hover { background: #d91d32 !important; }
.tool-menu {
position: fixed;
top: 20px;
right: 20px;
background: #fff;
border: 1px solid #ddd;
border-radius: 8px;
padding: 16px;
box-shadow: 0 4px 12px rgba(0,0,0,0.15);
z-index: 1000;
min-width: 200px;
}
.server-title {
font-weight: bold;
margin-bottom: 10px;
color: #333;
}
.expandable {
margin-top: 8px;
}
[data-testid="stSidebar"] .stSelectbox label {
color: #fff !important;
font-weight: 500;
font-size: 1.07rem;
}
/* Visualization styles */
.visualization-container {
margin: 20px 0;
padding: 15px;
border: 1px solid #ddd;
border-radius: 8px;
background: #f9f9f9;
}
.visualization-title {
font-size: 1.2rem;
font-weight: bold;
margin-bottom: 10px;
color: #333;
}
</style>
""", unsafe_allow_html=True)
# ========== DYNAMIC TOOL DISCOVERY FUNCTIONS ==========
async def _discover_tools() -> dict:
"""Discover available tools from the MCP server"""
try:
transport = StreamableHttpTransport(f"{st.session_state.get('MCP_SERVER_URL', 'http://localhost:8000')}/mcp")
async with Client(transport) as client:
tools = await client.list_tools()
return {tool.name: tool.description for tool in tools}
except Exception as e:
st.error(f"Failed to discover tools: {e}")
return {}
def discover_tools() -> dict:
"""Synchronous wrapper for tool discovery"""
return asyncio.run(_discover_tools())
def generate_tool_descriptions(tools_dict: dict) -> str:
"""Generate tool descriptions string from discovered tools"""
if not tools_dict:
return "No tools available"
descriptions = ["Available tools:"]
for i, (tool_name, tool_desc) in enumerate(tools_dict.items(), 1):
descriptions.append(f"{i}. {tool_name}: {tool_desc}")
return "\n".join(descriptions)
def get_image_base64(img_path):
img = Image.open(img_path)
buffered = BytesIO()
img.save(buffered, format="PNG")
img_bytes = buffered.getvalue()
img_base64 = base64.b64encode(img_bytes).decode()
return img_base64
# ========== SIDEBAR NAVIGATION ==========
with st.sidebar:
st.markdown("<div class='sidebar-title'>Solutions Scope</div>", unsafe_allow_html=True)
with st.container():
# Application selectbox (with key)
application = st.selectbox(
"Select Application",
["Select Application", "MCP Application"],
key="app_select"
)
# Dynamically choose default options for other selects
# Option lists
protocol_options = ["", "MCP Protocol", "A2A Protocol"]
llm_options = ["", "Groq Llama3-70B", "Groq Llama3-8B", "Groq Mixtral-8x7B", "Groq Gemma"]
# Logic to auto-select defaults if MCP Application is chosen
protocol_index = protocol_options.index(
"MCP Protocol") if application == "MCP Application" else protocol_options.index(
st.session_state.get("protocol_select", ""))
llm_index = llm_options.index("Groq Llama3-70B") if application == "MCP Application" else llm_options.index(
st.session_state.get("llm_select", ""))
protocol = st.selectbox(
"Protocol",
protocol_options,
key="protocol_select",
index=protocol_index
)
llm_model = st.selectbox(
"LLM Models",
llm_options,
key="llm_select",
index=llm_index
)
# Dynamic server tools selection based on discovered tools
if application == "MCP Application" and "available_tools" in st.session_state and st.session_state.available_tools:
server_tools_options = [""] + list(st.session_state.available_tools.keys())
default_tool = list(st.session_state.available_tools.keys())[0] if st.session_state.available_tools else ""
server_tools_index = server_tools_options.index(default_tool) if default_tool else 0
else:
server_tools_options = ["", "sqlserver_crud", "postgresql_crud"] # Fallback
server_tools_index = 0
server_tools = st.selectbox(
"Server Tools",
server_tools_options,
key="server_tools",
index=server_tools_index
)
st.button("Clear/Reset", key="clear_button")
st.markdown('<div class="sidebar-logo-label">Build & Deployed on</div>', unsafe_allow_html=True)
logo_base64 = get_image_base64("llm.png")
st.markdown(
f"""
<div class="sidebar-logo-row">
<img src="https://media.licdn.com/dms/image/v2/D560BAQFIon13R1UG4g/company-logo_200_200/company-logo_200_200/0/1733990910443/llm_at_scale_logo?e=2147483647&v=beta&t=WtAgFOcGQuTS0aEIqZhNMzWraHwL6FU0z5EPyPrty04" title="Logo" style="width: 50px; height: 50px;">
<img src="https://cdn.jsdelivr.net/gh/devicons/devicon/icons/googlecloud/googlecloud-original.svg" title="Google Cloud" style="width: 50px; height: 50px;">
<img src="https://a0.awsstatic.com/libra-css/images/logos/aws_logo_smile_1200x630.png" title="AWS" style="width: 50px; height: 50px;">
<img src="https://upload.wikimedia.org/wikipedia/commons/a/a8/Microsoft_Azure_Logo.svg" title="Azure Cloud" style="width: 50px; height: 50px;">
</div>
""",
unsafe_allow_html=True
)
# ========== LOGO/HEADER FOR MAIN AREA ==========
logo_path = "llm.png"
logo_base64 = get_image_base64(logo_path) if os.path.exists(logo_path) else ""
if logo_base64:
st.markdown(
f"""
<div style='display: flex; flex-direction: column; align-items: center; margin-bottom:20px;'>
<img src='data:image/png;base64,{logo_base64}' width='220'>
</div>
""",
unsafe_allow_html=True
)
st.markdown(
"""
<div style="
display: flex;
flex-direction: column;
align-items: center;
margin-bottom: 18px;
padding: 10px 0 10px 0;
">
<span style="
font-size: 2.5rem;
font-weight: bold;
letter-spacing: -2px;
color: #222;
">
MCP-Driven Data Management With Natural Language
</span>
<span style="
font-size: 1.15rem;
color: #555;
margin-top: 0.35rem;
">
Agentic Approach: NO SQL, NO ETL, NO DATA WAREHOUSING, NO BI TOOL
</span>
<hr style="
width: 80%;
border: none;
height: 2px;
background: linear-gradient(90deg, transparent, #4286f4, transparent);
margin: 20px auto;
">
</div>
""",
unsafe_allow_html=True
)
# ========== SESSION STATE INIT ==========
if "messages" not in st.session_state:
st.session_state.messages = []
# Initialize available_tools if not exists
if "available_tools" not in st.session_state:
st.session_state.available_tools = {}
# Initialize MCP_SERVER_URL in session state
if "MCP_SERVER_URL" not in st.session_state:
st.session_state["MCP_SERVER_URL"] = os.getenv("MCP_SERVER_URL", "http://localhost:8000")
# Initialize tool_states dynamically based on discovered tools
if "tool_states" not in st.session_state:
st.session_state.tool_states = {}
if "show_menu" not in st.session_state:
st.session_state["show_menu"] = False
if "menu_expanded" not in st.session_state:
st.session_state["menu_expanded"] = True
if "chat_input_box" not in st.session_state:
st.session_state["chat_input_box"] = ""
# Initialize visualization state
if "visualizations" not in st.session_state:
st.session_state.visualizations = []
# ========== HELPER FUNCTIONS ==========
def _clean_json(raw: str) -> str:
fences = re.findall(r"``````", raw, re.DOTALL)
if fences:
return fences[0].strip()
# If no JSON code fence, try to find JSON-like content
json_match = re.search(r'\{.*\}', raw, re.DOTALL)
return json_match.group(0).strip() if json_match else raw.strip()
import requests
def call_mcp_tool(tool_name: str, operation: str, args: dict) -> dict:
"""
Synchronous helper that calls the MCP server REST endpoint for a tool.
Adjust URL/path depending on your FastMCP HTTP transport.
"""
url = st.session_state.get("MCP_SERVER_URL", "http://localhost:8000") + f"/call_tool"
payload = {"tool": tool_name, "operation": operation, "args": args}
try:
resp = requests.post(url, json=payload, timeout=15)
resp.raise_for_status()
return resp.json()
except Exception as e:
return {"sql": None, "result": f"❌ error calling MCP tool: {e}"}
# ========== PARAMETER VALIDATION FUNCTION ==========
def validate_and_clean_parameters(tool_name: str, args: dict) -> dict:
"""Validate and clean parameters for specific tools"""
if tool_name == "sales_crud":
# Define allowed parameters for sales_crud (with WHERE clause support)
allowed_params = {
'operation', 'customer_id', 'product_id', 'quantity',
'unit_price', 'total_amount', 'sale_id', 'new_quantity',
'table_name', 'display_format', 'customer_name',
'product_name', 'email', 'total_price',
'columns', # Column selection
'where_clause', # WHERE conditions
'filter_conditions', # Structured filters
'limit' # Row limit
}
# Clean args to only include allowed parameters
cleaned_args = {k: v for k, v in args.items() if k in allowed_params}
# Validate display_format values
if 'display_format' in cleaned_args:
valid_formats = [
'Data Format Conversion',
'Decimal Value Formatting',
'String Concatenation',
'Null Value Removal/Handling'
]
if cleaned_args['display_format'] not in valid_formats:
cleaned_args.pop('display_format', None)
# Clean up columns parameter
if 'columns' in cleaned_args:
if isinstance(cleaned_args['columns'], str) and cleaned_args['columns'].strip():
columns_str = cleaned_args['columns'].strip()
columns_list = [col.strip() for col in columns_str.split(',') if col.strip()]
cleaned_args['columns'] = ','.join(columns_list)
else:
cleaned_args.pop('columns', None)
# Validate WHERE clause
if 'where_clause' in cleaned_args:
if not isinstance(cleaned_args['where_clause'], str) or not cleaned_args['where_clause'].strip():
cleaned_args.pop('where_clause', None)
# Validate limit
if 'limit' in cleaned_args:
try:
limit_val = int(cleaned_args['limit'])
if limit_val <= 0 or limit_val > 1000: # Reasonable limits
cleaned_args.pop('limit', None)
else:
cleaned_args['limit'] = limit_val
except (ValueError, TypeError):
cleaned_args.pop('limit', None)
return cleaned_args
elif tool_name == "sqlserver_crud":
allowed_params = {
'operation', 'name', 'email', 'limit', 'customer_id',
'new_email', 'table_name'
}
return {k: v for k, v in args.items() if k in allowed_params}
elif tool_name == "postgresql_crud":
allowed_params = {
'operation', 'name', 'price', 'description', 'limit',
'product_id', 'new_price', 'table_name'
}
return {k: v for k, v in args.items() if k in allowed_params}
return args
# ========== NEW LLM RESPONSE GENERATOR ==========
def generate_llm_response(operation_result: dict, action: str, tool: str, user_query: str, history_limit: int = 10) -> str:
"""Generate LLM response based on operation result with context (includes chat history)."""
# collect last N messages from session (if available)
messages_for_llm = []
history = st.session_state.get("messages", [])[-history_limit:]
for m in history:
role = m.get("role", "user")
content = m.get("content", "")
# convert to System/Human/Assistant roles for your LLM client
if role == "assistant":
messages_for_llm.append(HumanMessage(content=f"(assistant) {content}"))
else:
messages_for_llm.append(HumanMessage(content=f"(user) {content}"))
system_prompt = (
"You are a helpful database assistant. Generate a brief, natural response "
"explaining what operation was performed and its result. Be conversational "
"and informative. Focus on the business context and user-friendly explanation."
)
user_prompt = f"""
User asked: "{user_query}"
Operation: {action}
Tool used: {tool}
Result: {json.dumps(operation_result, indent=2)}
Please respond naturally and reference prior conversation context where helpful.
"""
try:
messages = [SystemMessage(content=system_prompt)] + messages_for_llm + [HumanMessage(content=user_prompt)]
response = groq_client.invoke(messages)
return response.content.strip()
except Exception as e:
# Fallback response if LLM call fails
if action == "read":
return f"Successfully retrieved data from {tool}."
elif action == "create":
return f"Successfully created new record in {tool}."
elif action == "update":
return f"Successfully updated record in {tool}."
elif action == "delete":
return f"Successfully deleted record from {tool}."
elif action == "describe":
return f"Successfully retrieved table schema from {tool}."
else:
return f"Operation completed successfully using {tool}."
# ========== VISUALIZATION GENERATOR ==========
def generate_visualization(data: any, user_query: str, tool: str) -> tuple:
"""
Generate JavaScript visualization code based on data and query.
Streams code live while generating, then renders.
Returns tuple of (HTML/JS code for the visualization, raw code).
"""
# Prepare context for the LLM
context = {
"user_query": user_query,
"tool": tool,
"data_type": type(data).__name__,
"data_sample": data[:5] if isinstance(data, list) and len(data) > 0 else data
}
system_prompt = """
You are a JavaScript visualization expert. Generate interactive charts using Chart.js.
Analyze the data structure and user query to determine the most appropriate visualization. Make it aesthetic and informative.
RULES:
1. Return ONLY raw HTML and JavaScript code
2. Use Chart.js for visualizations (include CDN link)
3. Make it responsive but set fixed height for charts (max 400px)
4. Include appropriate titles and labels based on the user query
5. Handle both tabular data and simple results
6. No markdown, no explanations, just code
7. If data is complex, create multiple chart types (bar, line, pie) but limit to 2-5 charts
8. Use container div with fixed height and overflow: auto
9. Add 'chart-container' class to all chart containers
"""
user_prompt = f"""
Create an interactive visualization for this data:
User Query: "{user_query}"
Tool Used: {tool}
Data Type: {context['data_type']}
Data Sample: {json.dumps(context['data_sample'], indent=2)}
Generate a comprehensive visualization that helps understand the data.
Focus on the most important insights from the query.
Make sure charts have fixed heights and don't overflow.
"""
try:
# Prepare messages
messages = [
SystemMessage(content=system_prompt),
HumanMessage(content=user_prompt)
]
# Placeholder to show live code generation
placeholder = st.empty()
code_accum = ""
# Stream response tokens
for event in groq_client.stream(messages):
token = getattr(event, "content", "")
if token:
code_accum += token
placeholder.code(code_accum, language="html")
visualization_code = code_accum.strip()
# Return both the code and the rendered HTML
return visualization_code, visualization_code
except Exception as e:
# Fallback to a simple table if visualization generation fails
if isinstance(data, list) and len(data) > 0:
fallback_code = f"""
<div class="visualization-container" style="height: 400px; overflow: auto;">
<div class="visualization-title">Data Table</div>
<div id="table-container"></div>
</div>
<script>
const data = {json.dumps(data)};
let tableHtml = '<table border="1" style="width:100%; border-collapse: collapse;">';
// Add headers
tableHtml += '<tr>';
Object.keys(data[0]).forEach(key => {{
tableHtml += `<th style="padding: 8px; background: #f2f2f2;">${{key}}</th>`;
}});
tableHtml += '</tr>';
// Add rows
data.forEach(row => {{
tableHtml += '<tr>';
Object.values(row).forEach(value => {{
tableHtml += `<td style="padding: 8px;">${{value}}</td>`;
}});
tableHtml += '</tr>';
}});
tableHtml += '</table>';
document.getElementById('table-container').innerHTML = tableHtml;
</script>
"""
else:
fallback_code = f"""
<div class="visualization-container" style="height: 200px; overflow: auto;">
<div class="visualization-title">Result</div>
<p>{str(data)}</p>
</div>
"""
return fallback_code, fallback_code
# Add this CSS for the split layout
st.markdown("""
<style>
.split-container {
display: flex;
width: 100%;
gap: 20px;
margin: 20px 0;
}
.code-panel {
flex: 1;
background: #f8f9fa;
border-radius: 8px;
padding: 15px;
border: 1px solid #e9ecef;
max-height: 500px;
overflow-y: auto;
}
.viz-panel {
flex: 1;
background: #f8f9fa;
border-radius: 8px;
padding: 15px;
border: 1px solid #e9ecef;
max-height: 500px;
overflow-y: auto;
}
.code-header, .viz-header {
font-weight: bold;
margin-bottom: 10px;
color: #333;
display: flex;
justify-content: space-between;
align-items: center;
}
.copy-button {
background: #4286f4;
color: white;
border: none;
padding: 5px 10px;
border-radius: 4px;
cursor: pointer;
font-size: 0.8rem;
}
.copy-button:hover {
background: #397dd2;
}
.chart-container {
height: 350px !important;
margin-bottom: 20px;
}
.visualization-container {
height: 400px;
overflow: auto;
}
</style>
""", unsafe_allow_html=True)
def parse_user_query(query: str, available_tools: dict) -> dict:
"""Enhanced parse user query with better DELETE operation handling"""
if not available_tools:
return {"error": "No tools available"}
# Build comprehensive tool information for the LLM
tool_info = []
for tool_name, tool_desc in available_tools.items():
tool_info.append(f"- **{tool_name}**: {tool_desc}")
tools_description = "\n".join(tool_info)
system_prompt = (
"You are an intelligent database router for CRUD operations. "
"Your job is to analyze the user's query and select the most appropriate tool based on the context and data being requested.\n\n"
"RESPONSE FORMAT:\n"
"Reply with exactly one JSON object: {\"tool\": string, \"action\": string, \"args\": object}\n\n"
"ACTION MAPPING:\n"
"- 'read': for viewing, listing, showing, displaying, or getting records\n"
"- 'create': for adding, inserting, or creating NEW records\n"
"- 'update': for modifying, changing, or updating existing records\n"
"- 'delete': for removing, deleting, or destroying records\n"
"- 'describe': for showing table structure, schema, or column information\n"
"- 'analyze': for analytical queries and statistical reports (calllogs_crud only)\n\n"
"CRITICAL TOOL SELECTION RULES:\n"
"\n"
"1. **PRODUCT QUERIES** → Use 'postgresql_crud':\n"
" - 'list products', 'show products', 'display products'\n"
" - 'product inventory', 'product catalog', 'product information'\n"
" - 'add product', 'create product', 'new product'\n"
" - 'update product', 'change product price', 'modify product'\n"
" - 'delete product', 'remove product', 'delete [ProductName]'\n"
" - Any query primarily about products, pricing, or inventory\n"
"\n"
"2. **CUSTOMER QUERIES** → Use 'sqlserver_crud':\n"
" - 'list customers', 'show customers', 'display customers'\n"
" - 'customer information', 'customer details'\n"
" - 'add customer', 'create customer', 'new customer'\n"
" - 'update customer', 'change customer email', 'modify customer'\n"
" - 'delete customer', 'remove customer', 'delete [CustomerName]'\n"
" - Any query primarily about customers, names, or emails\n"
"\n"
"3. **SALES/TRANSACTION QUERIES** → Use 'sales_crud':\n"
" - 'list sales', 'show sales', 'sales data', 'transactions'\n"
" - 'sales report', 'revenue data', 'purchase history'\n"
" - 'who bought what', 'customer purchases'\n"
" - Cross-database queries combining customer + product + sales info\n"
" - 'create sale', 'add sale', 'new transaction'\n"
" - Any query asking for combined data from multiple tables\n"
"\n"
"4. **CARE PLAN QUERIES** → Use 'careplan_crud':\n"
" - 'show care plans', 'list patients', 'display care plans', 'patient records'\n"
" - 'list care plans with name John', 'patients with diabetes'\n"
" - 'show care plan details', 'display patient information'\n"
" - 'patients needing housing assistance', 'care plans with employment status'\n"
" - 'reentry care plans', 'general care plans'\n"
" - Any query related to healthcare records, patient information, or care management\n\n"
"5. **CALL LOG ANALYSIS QUERIES** → Use 'calllogs_crud':\n"
" - 'analyze call logs', 'show call statistics', 'call center metrics'\n"
" - 'agent performance', 'sentiment analysis', 'issue frequency'\n"
" - 'call volume trends', 'escalation analysis', 'resolution rates'\n"
" - 'show calls by agent [AgentName]', 'calls with negative sentiment'\n"
" - 'call duration analysis', 'wait time statistics'\n"
" - 'top issue categories', 'service quality metrics'\n"
" - Use 'operation': 'analyze' for analytical reports\n"
" - Use 'operation': 'read' for raw call log data\n"
" - Any query related to call logs, agent performance, or customer service metrics\n\n"
"**ENHANCED CARE PLAN FIELD MAPPING:**\n"
"The CarePlan table now includes comprehensive real-world fields:\n"
"- Base: 'actual_release_date', 'name_of_youth', 'race_ethnicity', 'medi_cal_id_number'\n"
"- Health: 'health_screenings', 'health_assessments', 'chronic_conditions', 'prescribed_medications'\n"
"- Reentry: 'housing', 'employment', 'income_benefits', 'transportation', 'life_skills'\n"
"- Clinical: 'screenings', 'clinical_assessments', 'treatment_history', 'scheduled_appointments'\n"
"- Support: 'family_children', 'emergency_contacts', 'service_referrals', 'court_dates'\n"
"- Equipment: 'home_modifications', 'durable_medical_equipment'\n"
"- Metadata: 'care_plan_type', 'status', 'notes'\n\n"
"**ETL & DISPLAY FORMATTING RULES:**\n"
"For any data formatting requests (e.g., rounding decimals, changing date formats, handling nulls), "
"you MUST use the `display_format` parameter within the `sales_crud` tool.\n\n"
"1. **DECIMAL FORMATTING:**\n"
" - If the user asks to 'round', 'format to N decimal places', or mentions 'decimals'.\n"
" - Use: {\"display_format\": \"Decimal Value Formatting\"}\n"
" - **Example Query:** 'show sales with 2 decimal places'\n"
" - **→ Correct Tool Call:** {\"tool\": \"sales_crud\", \"action\": \"read\", \"args\": {\"display_format\": \"Decimal Value Formatting\"}}\n"
"2. **DATE FORMATTING:**\n"
" - If the user asks to 'format date', 'show date as YYYY-MM-DD', or similar.\n"
" - Use: {\"display_format\": \"Data Format Conversion\"}\n"
" - **Example Query:** 'show sales with formatted dates'\n"
" - **→ Correct Tool Call:** {\"tool\": \"sales_crud\", \"action\": \"read\", \"args\": {\"display_format\": \"Data Format Conversion\"}}\n"
"3. **NULL VALUE HANDLING:**\n"
" - If the user asks to 'remove nulls', 'replace empty values', or 'handle missing data'.\n"
" - Use: {\"display_format\": \"Null Value Removal/Handling\"}\n"
" - **Example Query:** 'show sales but remove records with missing info'\n"
" - **→ Correct Tool Call:** {\"tool\": \"sales_crud\", \"action\": \"read\", \"args\": {\"display_format\": \"Null Value Removal/Handling\"}}\n"
"4. **STRING CONCATENATION:**\n"
" - If the user asks to 'combine names', 'create a full description', or 'show full name'.\n"
" - Use: {\"display_format\": \"String Concatenation\"}\n"
" - **Example Query:** 'show sales with customer full names'\n"
" - **→ Correct Tool Call:** {\"tool\": \"sales_crud\", \"action\": \"read\", \"args\": {\"display_format\": \"String Concatenation\"}}\n"
"5. **CARE PLAN COLUMN FILTERING:**\n"
" - If the user asks to 'show only name and chronic conditions', 'remove address', or 'exclude phone'.\n"
" - Use: `columns` field in args with positive or negative column names.\n"
" - **Example Query:** 'show only name and chronic conditions from care plans'\n"
" - **→ Correct Tool Call:** {\"tool\": \"careplan_crud\", \"action\": \"read\", \"args\": {\"columns\": \"name_of_youth,chronic_conditions\"}}\n"
" - **Example Query:** 'show care plans without address and phone'\n"
" - **→ Correct Tool Call:** {\"tool\": \"careplan_crud\", \"action\": \"read\", \"args\": {\"columns\": \"*,-residential_address,-telephone\"}}\n"
"6. **CARE PLAN FILTERING BY TEXT OR VALUE:**\n"
" - If user asks 'care plans mentioning diabetes in chronic conditions', use LIKE\n"
" - Use: {\"where_clause\": \"chronic_conditions LIKE '%diabetes%'\"}\n"
" - **Example Query:** 'list patients with diabetes'\n"
" - **→ Correct Tool Call:** {\"tool\": \"careplan_crud\", \"action\": \"read\", \"args\": {\"where_clause\": \"chronic_conditions LIKE '%diabetes%'\"}}\n"
" - **Example Query:** 'care plans where name is John'\n"
" - **→ Correct Tool Call:** {\"tool\": \"careplan_crud\", \"action\": \"read\", \"args\": {\"where_clause\": \"name_of_youth = 'John'\"}}\n"
" - **Example Query:** 'show patients needing housing assistance'\n"
" - **→ Correct Tool Call:** {\"tool\": \"careplan_crud\", \"action\": \"read\", \"args\": {\"where_clause\": \"housing LIKE '%assistance%' OR housing = 'Homeless'\"}}\n"
"7. **CARE PLAN TYPE FILTERING:**\n"
" - If user asks for 'reentry care plans' or 'general care plans'\n"
" - Use: {\"care_plan_type\": \"Reentry Care Plan\"} or {\"care_plan_type\": \"General Care Plan\"}\n"
" - **Example Query:** 'show reentry care plans'\n"
" - **→ Correct Tool Call:** {\"tool\": \"careplan_crud\", \"action\": \"read\", \"args\": {\"care_plan_type\": \"Reentry Care Plan\"}}\n"
"8. **CALL LOGS ANALYSIS TYPES:**\n"
" - sentiment_by_agent: Agent sentiment performance\n"
" - issue_frequency: Most common issues\n"
" - call_volume_trends: Daily call trends\n"
" - escalation_analysis: Escalation rates by issue\n"
" - agent_performance: Comprehensive agent metrics\n"
" - **Example Query:** 'analyze agent performance'\n"
" - **→ Correct Tool Call:** {\"tool\": \"calllogs_crud\", \"action\": \"analyze\", \"args\": {\"analysis_type\": \"agent_performance\"}}\n"
"9. **CALL LOGS FILTERING:**\n"
" - Use 'agent_name', 'issue_category', 'sentiment_threshold' for filtering\n"
" - **Example Query:** 'show calls with negative sentiment'\n"
" - **→ Correct Tool Call:** {\"tool\": \"calllogs_crud\", \"action\": \"read\", \"args\": {\"sentiment_threshold\": -0.1}}\n"
" - **Example Query:** 'calls handled by Sarah Chen'\n"
" - **→ Correct Tool Call:** {\"tool\": \"calllogs_crud\", \"action\": \"read\", \"args\": {\"agent_name\": \"Sarah Chen\"}}\n"
)
user_prompt = f"""User query: "{query}"
Analyze the query step by step:
1. What is the PRIMARY INTENT? (product, customer, or sales operation)
2. What ACTION is being requested? (create, read, update, delete, describe)
3. What ENTITY NAME needs to be extracted? (for delete/update operations)
4. What SPECIFIC COLUMNS are requested? (for read operations - extract into 'columns' parameter)
5. What FILTER CONDITIONS are specified? (for read operations - extract into 'where_clause' parameter)
6. What PARAMETERS need to be extracted from the natural language?
ENTITY NAME EXTRACTION GUIDELINES (CRITICAL FOR DELETE/UPDATE):
- For "delete Widget" → extract "Widget" and put in 'name' parameter
- For "delete product Gadget" → extract "Gadget" and put in 'name' parameter
- For "delete customer Alice" → extract "Alice" and put in 'name' parameter
- For "update price of Tool to 30" → extract "Tool" and put in 'name' parameter, extract "30" and put in 'new_price'
COLUMN EXTRACTION GUIDELINES:
- Look for patterns like "show X, Y", "display X and Y", "get X, Y from Z"
- Extract only the column names, map them to standard names
- Put them in a comma-separated string in the 'columns' parameter
WHERE CLAUSE EXTRACTION GUIDELINES:
- Look for filtering conditions like "exceed", "above", "greater than", "with price over"
- Convert natural language to SQL-like conditions
- Handle currency symbols and numbers properly
- Put the condition in the 'where_clause' parameter
Respond with the exact JSON format with properly extracted parameters."""
try:
messages = [
SystemMessage(content=system_prompt),
HumanMessage(content=user_prompt)
]
resp = groq_client.invoke(messages)
raw = _clean_json(resp.content)
try:
result = json.loads(raw)
except json.JSONDecodeError:
try:
result = ast.literal_eval(raw)
except:
result = {"tool": list(available_tools.keys())[0], "action": "read", "args": {}}
# Normalize action names
if "action" in result and result["action"] in ["list", "show", "display", "view", "get"]:
result["action"] = "read"
# ENHANCED parameter extraction for DELETE and UPDATE operations
if result.get("action") in ["delete", "update"]:
args = result.get("args", {})
# Extract entity name for delete/update operations if not already extracted
if "name" not in args:
import re
# Enhanced regex patterns for delete operations
delete_patterns = [
r'(?:delete|remove)\s+customer\s+([A-Za-z]+(?:\s+[A-Za-z]+)?)',
r'(?:delete|remove)\s+product\s+([A-Za-z]+(?:\s+[A-Za-z]+)?)',
r'(?:delete|remove)\s+([A-Za-z]+(?:\s+[A-Za-z]+)?)',
r'(?:update|change)\s+(?:price\s+of\s+)?([A-Za-z]+(?:\s+[A-Za-z]+)?)',
]
for pattern in delete_patterns:
match = re.search(pattern, query, re.IGNORECASE)
if match:
extracted_name = match.group(1).strip()
# Clean up common words that might be captured
stop_words = ['product', 'customer', 'price', 'email', 'to', 'of', 'the', 'a', 'an']
name_words = [word for word in extracted_name.split() if word.lower() not in stop_words]
if name_words:
args["name"] = ' '.join(name_words)
break
# Extract new_price for product updates
if result.get("action") == "update" and result.get("tool") == "postgresql_crud" and "new_price" not in args:
import re
price_match = re.search(r'(?:to|=|\s+)\$?(\d+(?:\.\d+)?)', query, re.IGNORECASE)
if price_match:
args["new_price"] = float(price_match.group(1))
# Extract new_email for customer updates
if result.get("action") == "update" and result.get("tool") == "sqlserver_crud" and "new_email" not in args:
import re
email_match = re.search(r'(?:to|=|\s+)([\w\.-]+@[\w\.-]+\.\w+)', query, re.IGNORECASE)
if email_match:
args["new_email"] = email_match.group(1)
result["args"] = args
# Enhanced parameter extraction for create operations
elif result.get("action") == "create":
args = result.get("args", {})
# Extract name and email from query if not already extracted
if result.get("tool") == "sqlserver_crud" and ("name" not in args or "email" not in args):
# Try to extract name and email using regex patterns
import re
# Extract email
email_match = re.search(r'[\w\.-]+@[\w\.-]+\.\w+', query)
if email_match and "email" not in args:
args["email"] = email_match.group(0)
# Extract name (everything between 'customer' and 'with' or before email)
if "name" not in args:
# Pattern 1: "create customer [Name] with [email]"
name_match = re.search(r'(?:create|add|new)\s+customer\s+([^@]+?)(?:\s+with|\s+[\w\.-]+@)', query, re.IGNORECASE)
if not name_match:
# Pattern 2: "create [Name] [email]" or "add [Name] with [email]"
name_match = re.search(r'(?:create|add|new)\s+([^@]+?)(?:\s+with|\s+[\w\.-]+@)', query, re.IGNORECASE)
if not name_match:
# Pattern 3: Extract everything before the email
if email_match:
name_part = query[:email_match.start()].strip()
name_match = re.search(r'(?:customer|create|add|new)\s+(.+)', name_part, re.IGNORECASE)
if name_match:
extracted_name = name_match.group(1).strip()
# Clean up common words
extracted_name = re.sub(r'\b(with|email|named|called)\b', '', extracted_name, flags=re.IGNORECASE).strip()
if extracted_name:
args["name"] = extracted_name
result["args"] = args
# Enhanced parameter extraction for read operations with columns and where_clause
elif result.get("action") == "read" and result.get("tool") == "sales_crud":
args = result.get("args", {})
# Extract columns if not already extracted
if "columns" not in args:
import re
# Look for column specification patterns
column_patterns = [
r'(?:show|display|get|select)\s+([^,\s]+(?:,\s*[^,\s]+)*?)(?:\s+from|\s+where|\s*$)',
r'(?:show|display|get|select)\s+(.+?)\s+(?:from|where)',
r'display\s+(.+?)(?:\s+from|\s*$)',
]
for pattern in column_patterns:
match = re.search(pattern, query, re.IGNORECASE)
if match:
columns_text = match.group(1).strip()
# Clean up and standardize column names
if 'and' in columns_text or ',' in columns_text:
# Multiple columns
columns_list = re.split(r'[,\s]+and\s+|,\s*', columns_text)
cleaned_columns = []
for col in columns_list:
col = col.strip().lower().replace(' ', '_')
# Map common variations
if col in ['name', 'customer']:
cleaned_columns.append('customer_name')
elif col in ['price', 'total', 'amount']:
cleaned_columns.append('total_price')
elif col in ['product']:
cleaned_columns.append('product_name')
elif col in ['date']:
cleaned_columns.append('sale_date')
elif col in ['email']:
cleaned_columns.append('customer_email')
else:
cleaned_columns.append(col)
if cleaned_columns:
args["columns"] = ','.join(cleaned_columns)
else:
# Single column
col = columns_text.strip().lower().replace(' ', '_')
if col in ['name', 'customer']:
args["columns"] = 'customer_name'
elif col in ['price', 'total', 'amount']:
args["columns"] = 'total_price'
elif col in ['product']:
args["columns"] = 'product_name'
elif col in ['date']:
args["columns"] = 'sale_date'
elif col in ['email']:
args["columns"] = 'customer_email'
else:
args["columns"] = col
break
# Extract where_clause if not already extracted
if "where_clause" not in args:
import re
# Look for filtering conditions
where_patterns = [
r'(?:with|where)\s+total[_\s]*price[_\s]*(?:exceed[s]?|above|greater\s+than|more\s+than|>)\s*\$?(\d+(?:\.\d+)?)',
r'(?:with|where)\s+total[_\s]*price[_\s]*(?:below|less\s+than|under|<)\s*\$?(\d+(?:\.\d+)?)',
r'(?:with|where)\s+total[_\s]*price[_\s]*(?:equal[s]?|is|=)\s*\$?(\d+(?:\.\d+)?)',
r'(?:with|where)\s+quantity[_\s]*(?:>|above|greater\s+than|more\s+than)\s*(\d+)',
r'(?:with|where)\s+quantity[_\s]*(?:<|below|less\s+than|under)\s*(\d+)',
r'(?:with|where)\s+quantity[_\s]*(?:=|equal[s]?|is)\s*(\d+)',
r'(?:by|for)\s+customer[_\s]*([A-Za-z\s]+?)(?:\s|$)',
r'(?:for|of)\s+product[_\s]*([A-Za-z\s]+?)(?:\s|$)',
]
for i, pattern in enumerate(where_patterns):
match = re.search(pattern, query, re.IGNORECASE)
if match:
value = match.group(1).strip()
if i <= 2: # total_price conditions
if 'exceed' in query.lower() or 'above' in query.lower() or 'greater' in query.lower() or 'more' in query.lower():
args["where_clause"] = f"total_price > {value}"
elif 'below' in query.lower() or 'less' in query.lower() or 'under' in query.lower():
args["where_clause"] = f"total_price < {value}"
else:
args["where_clause"] = f"total_price = {value}"
elif i <= 5: # quantity conditions
if 'above' in query.lower() or 'greater' in query.lower() or 'more' in query.lower():
args["where_clause"] = f"quantity > {value}"
elif 'below' in query.lower() or 'less' in query.lower() or 'under' in query.lower():
args["where_clause"] = f"quantity < {value}"
else:
args["where_clause"] = f"quantity = {value}"
elif i == 6: # customer name
args["where_clause"] = f"customer_name = '{value}'"
elif i == 7: # product name
args["where_clause"] = f"product_name = '{value}'"
break
result["args"] = args
# Validate and clean args
if "args" in result and isinstance(result["args"], dict):
cleaned_args = validate_and_clean_parameters(result.get("tool"), result["args"])
result["args"] = cleaned_args
# Validate tool selection
if "tool" in result and result["tool"] not in available_tools:
result["tool"] = list(available_tools.keys())[0]
return result
except Exception as e:
return {
"tool": list(available_tools.keys())[0] if available_tools else None,
"action": "read",
"args": {},
"error": f"Failed to parse query: {str(e)}"
}
async def _invoke_tool(tool: str, action: str, args: dict) -> any:
transport = StreamableHttpTransport(f"{st.session_state['MCP_SERVER_URL']}/mcp")
async with Client(transport) as client:
payload = {"operation": action, **{k: v for k, v in args.items() if k != "operation"}}
res_obj = await client.call_tool(tool, payload)
if res_obj.structured_content is not None:
return res_obj.structured_content
text = "".join(b.text for b in res_obj.content).strip()
if text.startswith("{") and "}{" in text:
text = "[" + text.replace("}{", "},{") + "]"
try:
return json.loads(text)
except:
return text
def call_mcp_tool(tool: str, action: str, args: dict) -> any:
return asyncio.run(_invoke_tool(tool, action, args))
def format_natural(data) -> str:
if isinstance(data, list):
lines = []
for i, item in enumerate(data, 1):
if isinstance(item, dict):
parts = [f"{k} {v}" for k, v in item.items()]
lines.append(f"Record {i}: " + ", ".join(parts) + ".")
else:
lines.append(f"{i}. {item}")
return "\n".join(lines)
if isinstance(data, dict):
parts = [f"{k} {v}" for k, v in data.items()]
return ", ".join(parts) + "."
return str(data)
def normalize_args(args):
mapping = {
"product_name": "name",
"customer_name": "name",
"item": "name"
}
for old_key, new_key in mapping.items():
if old_key in args:
args[new_key] = args.pop(old_key)
return args
def extract_name_from_query(text: str) -> str:
"""Enhanced name extraction that handles various patterns"""
# Patterns for customer operations
customer_patterns = [
r'delete\s+customer\s+([A-Za-z]+(?:\s+[A-Za-z]+)?)',
r'remove\s+customer\s+([A-Za-z]+(?:\s+[A-Za-z]+)?)',
r'update\s+customer\s+([A-Za-z]+(?:\s+[A-Za-z]+)?)',
r'delete\s+([A-Za-z]+(?:\s+[A-Za-z]+)?)',
r'remove\s+([A-Za-z]+(?:\s+[A-Za-z]+)?)'
]
# Patterns for product operations
product_patterns = [
r'delete\s+product\s+([A-Za-z]+(?:\s+[A-Za-z]+)?)',
r'remove\s+product\s+([A-Za-z]+(?:\s+[A-Za-z]+)?)',
r'update\s+(?:price\s+of\s+)?([A-Za-z]+(?:\s+[A-Za-z]+)?)',
r'change\s+price\s+of\s+([A-Za-z]+(?:\s+[A-Za-z]+)?)',
r'(?:price\s+of\s+)([A-Za-z]+(?:\s+[A-Za-z]+)?)\s+(?:to|=)'
]
all_patterns = customer_patterns + product_patterns
for pattern in all_patterns:
match = re.search(pattern, text, re.IGNORECASE)
if match:
return match.group(1).strip()
return None
def extract_email(text):
match = re.search(r'[\w\.-]+@[\w\.-]+', text)
return match.group(0) if match else None
def extract_price(text):
# Look for price patterns like "to 25", "= 30.50", "$15.99"
price_patterns = [
r'to\s+\$?(\d+(?:\.\d+)?)',
r'=\s+\$?(\d+(?:\.\d+)?)',
r'\$(\d+(?:\.\d+)?)',
r'(\d+(?:\.\d+)?)\s*dollars?'
]
for pattern in price_patterns:
match = re.search(pattern, text, re.IGNORECASE)
if match:
return float(match.group(1))
return None
def generate_table_description(df: pd.DataFrame, content: dict, action: str, tool: str) -> str:
"""Generate LLM-based table description from JSON response data"""
# Sample first few rows for context (don't send all data to LLM)
sample_data = df.head(3).to_dict('records') if len(df) > 0 else []
# Create context for LLM
context = {
"action": action,
"tool": tool,
"record_count": len(df),
"columns": list(df.columns) if len(df) > 0 else [],
"sample_data": sample_data,
"full_response": content.get("result", [])[:3] if isinstance(content.get("result"), list) else content.get(
"result", "")
}
system_prompt = (
"You are a data analyst. Generate a brief, insightful 1-line description "
"of the table data based on the JSON response. Focus on what the data represents "
"and any interesting patterns you notice. Be concise and business-focused."
)
user_prompt = f"""
Analyze this table data and generate a single insightful line about it:
Context: {json.dumps(context, indent=2)}
Generate one line describing what this data represents and any key insights.
"""
try:
messages = [
SystemMessage(content=system_prompt),
HumanMessage(content=user_prompt)
]
response = groq_client.invoke(messages)
return response.content.strip()
except Exception as e:
return f"Retrieved {len(df)} records from the database."
# ========== MAIN ==========
if application == "MCP Application":
user_avatar_url = "https://cdn-icons-png.flaticon.com/512/1946/1946429.png"
agent_avatar_url = "https://cdn-icons-png.flaticon.com/512/4712/4712039.png"
MCP_SERVER_URL = os.getenv("MCP_SERVER_URL", "http://localhost:8000")
st.session_state["MCP_SERVER_URL"] = MCP_SERVER_URL
# Discover tools dynamically if not already done
if not st.session_state.available_tools:
with st.spinner("Discovering available tools..."):
discovered_tools = discover_tools()
st.session_state.available_tools = discovered_tools
st.session_state.tool_states = {tool: True for tool in discovered_tools.keys()}
# Generate dynamic tool descriptions
TOOL_DESCRIPTIONS = generate_tool_descriptions(st.session_state.available_tools)
# ========== TOOLS STATUS AND REFRESH BUTTON ==========
# Create columns for tools info and refresh button
col1, col2 = st.columns([4, 1])
with col1:
# Display discovered tools info
if st.session_state.available_tools:
st.info(
f"🔧 Discovered {len(st.session_state.available_tools)} tools: {', '.join(st.session_state.available_tools.keys())}")
else:
st.warning("⚠️ No tools discovered. Please check your MCP server connection.")
with col2:
# Small refresh button on main page
st.markdown('<div class="small-refresh-button">', unsafe_allow_html=True)
if st.button("🔄 Active Server", key="refresh_tools_main", help="Rediscover available tools"):
with st.spinner("Refreshing tools..."):
MCP_SERVER_URL = os.getenv("MCP_SERVER_URL", "http://localhost:8000")
st.session_state["MCP_SERVER_URL"] = MCP_SERVER_URL
discovered_tools = discover_tools()
st.session_state.available_tools = discovered_tools
st.session_state.tool_states = {tool: True for tool in discovered_tools.keys()}
st.rerun()
st.markdown('</div>', unsafe_allow_html=True)
# ========== 1. RENDER CHAT MESSAGES ==========
st.markdown('<div class="stChatPaddingBottom">', unsafe_allow_html=True)
for msg in st.session_state.messages:
if msg["role"] == "user":
st.markdown(
f"""
<div class="chat-row right">
<div class="chat-bubble user-msg user-bubble">{msg['content']}</div>
<img src="{user_avatar_url}" class="avatar user-avatar" alt="User">
</div>
""",
unsafe_allow_html=True,
)
elif msg.get("format") == "reasoning":
st.markdown(
f"""
<div class="chat-row left">
<img src="{agent_avatar_url}" class="avatar agent-avatar" alt="Agent">
<div class="chat-bubble agent-msg agent-bubble"><i>{msg['content']}</i></div>
</div>
""",
unsafe_allow_html=True,
)
elif msg.get("format") == "multi_step_read" and isinstance(msg["content"], dict):
step = msg["content"]
st.markdown(
f"""
<div class="chat-row left">
<img src="{agent_avatar_url}" class="avatar agent-avatar" alt="Agent">
<div class="chat-bubble agent-msg agent-bubble">
<b>Step: Lookup by name</b> (<code>{step['args'].get('name', '')}</code>)
</div>
</div>
""", unsafe_allow_html=True
)
with st.expander(f"Lookup Request: {step['tool']}"):
st.code(json.dumps({
"tool": step['tool'],
"action": step['action'],
"args": step['args']
}, indent=2), language="json")
if isinstance(step["result"], dict) and "sql" in step["result"]:
with st.expander("Lookup SQL Query Used"):
st.code(step["result"]["sql"], language="sql")
if isinstance(step["result"], dict) and "result" in step["result"]:
with st.expander("Lookup Response"):
st.code(json.dumps(step["result"]["result"], indent=2), language="json")
if isinstance(step["result"]["result"], list) and step["result"]["result"]:
df = pd.DataFrame(step["result"]["result"])
st.markdown("**Lookup Result Table:**")
st.table(df)
elif msg.get("format") == "sql_crud" and isinstance(msg["content"], dict):
content = msg["content"]
action = msg.get("action", "")
tool = msg.get("tool", "")
user_query = msg.get("user_query", "")
with st.expander("Details"):
if "request" in msg:
st.markdown("**Request**")
st.code(json.dumps(msg["request"], indent=2), language="json")
st.markdown("---")
st.markdown("**SQL Query Used**")
st.code(content["sql"] or "No SQL executed", language="sql")
st.markdown("---")
st.markdown("**Response**")
if isinstance(content["result"], (dict, list)):
st.code(json.dumps(content["result"], indent=2), language="json")
else:
st.code(content["result"])
# Generate LLM response for the operation
llm_response = generate_llm_response(content, action, tool, user_query)
st.markdown(
f"""
<div class="chat-row left">
<img src="{agent_avatar_url}" class="avatar agent-avatar" alt="Agent">
<div class="chat-bubble agent-msg agent-bubble">{llm_response}</div>
</div>
""", unsafe_allow_html=True
)
if action in {"create", "update", "delete"}:
result_msg = content.get("result", "")
if "✅" in result_msg or "success" in result_msg.lower():
st.success(result_msg)
elif "❌" in result_msg or "fail" in result_msg.lower() or "error" in result_msg.lower():
st.error(result_msg)
else:
st.info(result_msg)
try:
st.markdown("#### Here's the updated table after your operation:")
read_tool = tool
read_args = {}
updated_table = call_mcp_tool(read_tool, "read", read_args)
if isinstance(updated_table, dict) and "result" in updated_table:
updated_df = pd.DataFrame(updated_table["result"])
st.table(updated_df)
except Exception as fetch_err:
st.info(f"Could not retrieve updated table: {fetch_err}")
if action == "read" and isinstance(content["result"], list):
st.markdown("#### Here's the current table:")
df = pd.DataFrame(content["result"])
st.table(df)
# Check if this is ETL formatted data by looking for specific formatting
if tool == "sales_crud" and len(df.columns) > 0:
# Check for different ETL formats based on column names
if "sale_summary" in df.columns:
st.info("📊 Data formatted with String Concatenation - Combined fields for readability")
elif "sale_date" in df.columns and isinstance(df["sale_date"].iloc[0] if len(df) > 0 else None,
str):
st.info("📅 Data formatted with Data Format Conversion - Dates converted to string format")
elif any(
"." in str(val) and len(str(val).split(".")[-1]) == 2 for val in df.get("unit_price", []) if
pd.notna(val)):
st.info("💰 Data formatted with Decimal Value Formatting - Prices formatted to 2 decimal places")
else:
st.markdown(f"The table contains {len(df)} sales records with cross-database information.")
elif tool == "sqlserver_crud":
st.markdown(
f"The table contains {len(df)} customers with their respective IDs, names, emails, and creation timestamps."
)
elif tool == "postgresql_crud":
st.markdown(
f"The table contains {len(df)} products with their respective IDs, names, prices, and descriptions."
)
else:
st.markdown(f"The table contains {len(df)} records.")
elif action == "describe" and isinstance(content['result'], list):
st.markdown("#### Table Schema: ")
df = pd.DataFrame(content['result'])
st.table(df)
st.markdown(
"This shows the column names, data types, nullability, and maximum length for each column in the table.")
else:
st.markdown(
f"""
<div class="chat-row left">
<img src="{agent_avatar_url}" class="avatar agent-avatar" alt="Agent">
<div class="chat-bubble agent-msg agent-bubble">{msg['content']}</div>
</div>
""",
unsafe_allow_html=True,
)
# Render saved visualizations (most recent first)
# Auto-render saved visualizations (newest first)
if st.session_state.get("visualizations"):
st.markdown("### Visualizations")
for idx, (viz_html, viz_code, q) in enumerate(reversed(st.session_state["visualizations"])):
# First (newest) one stays expanded, others collapsed
expanded_state = True if idx == 0 else False
with st.expander(f"📊 Visualization for: {q}", expanded=expanded_state):
st.markdown("**Generated Code:**")
import time
if expanded_state:
placeholder = st.empty()
typed = ""
for char in viz_code:
typed += char
placeholder.code(typed, language="html")
time.sleep(0.01) # adjust speed (0.005 faster, 0.05 slower)
else:
st.code(viz_code, language="html")
st.markdown("**Rendered Chart:**")
try:
components.html(viz_html, height=420, scrolling=True)
except Exception as e:
st.error(f"Error rendering visualization: {e}")
st.markdown('</div>', unsafe_allow_html=True) # End stChatPaddingBottom
# ========== 2. RENDER VISUALIZATIONS ==========
if st.session_state.visualizations:
st.markdown("---")
st.markdown("## 📊 Interactive Visualizations")
for i, (viz_html, viz_code, user_query) in enumerate(st.session_state.visualizations):
with st.expander(
f"Visualization: {user_query[:50]}..." if len(user_query) > 50 else f"Visualization: {user_query}",expanded=False):
# Create tabs with Code first, then Visualization
tab1, tab2 = st.tabs(["💻 Generated Code", "📊 Visualization"])
with tab1:
st.markdown("**Generated Code**")
# Initialize streaming state for this visualization if not exists
stream_key = f"stream_complete_{i}"
if stream_key not in st.session_state:
st.session_state[stream_key] = False
# Create placeholder for streaming effect
code_placeholder = st.empty()
if not st.session_state[stream_key]:
# Streaming effect - show code character by character
import time
# Show streaming indicator first
with code_placeholder.container():
st.info("🔄 Generating code...")
# Small delay to show the loading message
time.sleep(0.5)
# Stream the code
streamed_code = ""
for j, char in enumerate(viz_code):
streamed_code += char
# Update every 5-10 characters for better performance
if j % 8 == 0 or j == len(viz_code) - 1:
code_placeholder.code(streamed_code, language="html")
time.sleep(0.03) # Adjust speed as needed
# Mark streaming as complete
st.session_state[stream_key] = True
# Force a rerun to show the complete state
st.rerun()
else:
# Show complete code immediately
code_placeholder.code(viz_code, language="html")
# Adding copy button (only show when streaming is complete)
if st.session_state[stream_key]:
if st.button("📋 Copy Code", key=f"copy_{i}"):
st.session_state.copied_code = viz_code
st.success("Code copied to clipboard!")
# Add reset streaming button for demo purposes
if st.button("🔄 Replay Code Generation", key=f"replay_{i}"):
st.session_state[stream_key] = False
st.rerun()
with tab2:
st.markdown("**Interactive Visualization**")
# Use a container with fixed height
with st.container():
components.html(viz_html, height=400, scrolling=True)
if st.button("🧹 Clear All Visualizations", key="clear_viz"):
st.session_state.visualizations = []
# Clear all streaming states
keys_to_remove = [key for key in st.session_state.keys() if key.startswith("stream_complete_")]
for key in keys_to_remove:
del st.session_state[key]
st.rerun()
# ========== 3. CLAUDE-STYLE STICKY CHAT BAR ==========
st.markdown('<div class="sticky-chatbar"><div class="chatbar-claude">', unsafe_allow_html=True)
with st.form("chatbar_form", clear_on_submit=True):
chatbar_cols = st.columns([1, 16, 1]) # Left: hamburger, Middle: input, Right: send
# --- LEFT: Hamburger (Tools) ---
with chatbar_cols[0]:
hamburger_clicked = st.form_submit_button("≡", use_container_width=True)
# --- MIDDLE: Input Box ---
with chatbar_cols[1]:
user_query_input = st.text_input(
"Chat Input", # Provide a label
placeholder="How can I help you today?",
label_visibility="collapsed", # Hide the label visually
key="chat_input_box"
)
# --- RIGHT: Send Button ---
with chatbar_cols[2]:
send_clicked = st.form_submit_button("➤", use_container_width=True)
st.markdown('</div></div>', unsafe_allow_html=True)
# ========== FLOATING TOOL MENU ==========
if st.session_state.get("show_menu", False):
st.markdown('<div class="tool-menu">', unsafe_allow_html=True)
st.markdown('<div class="server-title">MultiDBCRUD</div>', unsafe_allow_html=True)
tool_label = "Tools" + (" ▼" if st.session_state["menu_expanded"] else " ▶")
if st.button(tool_label, key="expand_tools", help="Show tools", use_container_width=True):
st.session_state["menu_expanded"] = not st.session_state["menu_expanded"]
if st.session_state["menu_expanded"]:
st.markdown('<div class="expandable">', unsafe_allow_html=True)
for tool in st.session_state.tool_states.keys():
enabled = st.session_state.tool_states[tool]
new_val = st.toggle(tool, value=enabled, key=f"tool_toggle_{tool}")
st.session_state.tool_states[tool] = new_val
st.markdown('</div>', unsafe_allow_html=True)
st.markdown('</div>', unsafe_allow_html=True)
# ========== HANDLE HAMBURGER ==========
if hamburger_clicked:
st.session_state["show_menu"] = not st.session_state.get("show_menu", False)
st.rerun()
# ========== PROCESS CHAT INPUT ==========
if user_query_input and send_clicked:
user_query = user_query_input
user_steps = []
try:
enabled_tools = [k for k, v in st.session_state.tool_states.items() if v]
if not enabled_tools:
raise Exception("No tools are enabled. Please enable at least one tool in the menu.")
p = parse_user_query(user_query, st.session_state.available_tools)
tool = p.get("tool")
if tool not in enabled_tools:
raise Exception(f"Tool '{tool}' is disabled. Please enable it in the menu.")
if tool not in st.session_state.available_tools:
raise Exception(
f"Tool '{tool}' is not available. Available tools: {', '.join(st.session_state.available_tools.keys())}")
action = p.get("action")
args = p.get("args", {})
# VALIDATE AND CLEAN PARAMETERS
args = validate_and_clean_parameters(tool, args)
args = normalize_args(args)
p["args"] = args
# ========== ENHANCED NAME-BASED RESOLUTION ==========
# For SQL Server (customers) operations
if tool == "sqlserver_crud":
if action in ["update", "delete"] and "name" in args and "customer_id" not in args:
# First, try to find the customer by name
name_to_find = args["name"]
try:
# Search for customer by name
read_result = call_mcp_tool(tool, "read", {})
if isinstance(read_result, dict) and "result" in read_result:
customers = read_result["result"]
# Try exact match first
exact_matches = [c for c in customers if c.get("Name", "").lower() == name_to_find.lower()]
if exact_matches:
args["customer_id"] = exact_matches[0]["Id"]
else:
# Try partial matches (first name or last name)
partial_matches = [c for c in customers if
name_to_find.lower() in c.get("Name", "").lower() or
name_to_find.lower() in c.get("FirstName", "").lower() or
name_to_find.lower() in c.get("LastName", "").lower()]
if partial_matches:
args["customer_id"] = partial_matches[0]["Id"]
else:
raise Exception(f"❌ Customer '{name_to_find}' not found")
except Exception as e:
if "not found" in str(e):
raise e
else:
raise Exception(f"❌ Error finding customer '{name_to_find}': {str(e)}")
# Extract new email for updates
if action == "update" and "new_email" not in args:
possible_email = extract_email(user_query)
if possible_email:
args["new_email"] = possible_email
# For PostgreSQL (products) operations
elif tool == "postgresql_crud":
if action in ["update", "delete"] and "name" in args and "product_id" not in args:
# First, try to find the product by name
name_to_find = args["name"]
try:
# Search for product by name
read_result = call_mcp_tool(tool, "read", {})
if isinstance(read_result, dict) and "result" in read_result:
products = read_result["result"]
# Try exact match first
exact_matches = [p for p in products if p.get("name", "").lower() == name_to_find.lower()]
if exact_matches:
args["product_id"] = exact_matches[0]["id"]
else:
# Try partial matches
partial_matches = [p for p in products if name_to_find.lower() in p.get("name", "").lower()]
if partial_matches:
args["product_id"] = partial_matches[0]["id"]
else:
raise Exception(f"❌ Product '{name_to_find}' not found")
except Exception as e:
if "not found" in str(e):
raise e
else:
raise Exception(f"❌ Error finding product '{name_to_find}': {str(e)}")
# Extract new price for updates
if action == "update" and "new_price" not in args:
possible_price = extract_price(user_query)
if possible_price is not None:
args['new_price'] = possible_price
# Update the parsed args
p["args"] = args
# Handle describe operations
if action == "describe" and "table_name" in args:
if tool == "sqlserver_crud" and args["table_name"].lower() in ["customer", "customer table"]:
args["table_name"] = "Customers"
if tool == "postgresql_crud" and args["table_name"].lower() in ["product", "product table"]:
args["table_name"] = "products"
raw = call_mcp_tool(p["tool"], p["action"], p.get("args", {}))
# ========== GENERATE VISUALIZATION ==========
# Extract data for visualization
viz_data = raw
if isinstance(raw, dict) and "result" in raw:
viz_data = raw["result"]
# Generate visualization for read operations with data
if action == "read" and viz_data and (
(isinstance(viz_data, list) and len(viz_data) > 0) or
(isinstance(viz_data, dict) and len(viz_data) > 0)
):
with st.spinner("Generating visualization..."):
viz_code, viz_html = generate_visualization(viz_data, user_query, tool)
# Add to visualization list with both code and HTML
if "visualizations" not in st.session_state:
st.session_state.visualizations = []
st.session_state.visualizations.append((viz_html, viz_code, user_query))
st.success("Visualization generated successfully!")
except Exception as e:
reply, fmt = f"⚠️ Error: {e}", "text"
assistant_message = {
"role": "assistant",
"content": reply,
"format": fmt,
}
st.session_state.messages.append({
"role": "user",
"content": user_query,
"format": "text",
})
st.session_state.messages.append(assistant_message)
else:
st.session_state.messages.append({
"role": "user",
"content": user_query,
"format": "text",
})
for step in user_steps:
st.session_state.messages.append(step)
if isinstance(raw, dict) and "sql" in raw and "result" in raw:
reply, fmt = raw, "sql_crud"
else:
reply, fmt = format_natural(raw), "text"
assistant_message = {
"role": "assistant",
"content": reply,
"format": fmt,
"request": p,
"tool": p.get("tool"),
"action": p.get("action"),
"args": p.get("args"),
"user_query": user_query, # Added user_query to the message
}
st.session_state.messages.append(assistant_message)
st.rerun() # Rerun so chat output appears
# ========== 4. AUTO-SCROLL TO BOTTOM ==========
components.html("""
<script>
setTimeout(() => { window.scrollTo(0, document.body.scrollHeight); }, 80);
</script>
""")
# ========== ETL EXAMPLES HELP SECTION ==========
with st.expander("🔧 ETL Functions & Examples"):
st.markdown("""
### ETL Display Formatting Functions
Your MCP server supports 4 ETL (Extract, Transform, Load) functions for data formatting:
#### 1. Data Format Conversion
- **Query Examples:**
- "show sales with data format conversion"
- "convert sales data format"
- "format sales data for export"
- **What it does:** Converts dates to string format, removes unnecessary fields
#### 2. Decimal Value Formatting
- **Query Examples:**
- "format sales prices with decimal formatting"
- "show sales with 2 decimal places"
- "decimal value formatting for sales"
- **What it does:** Formats all prices to exactly 2 decimal places as strings
#### 3. String Concatenation
- **Query Examples:**
- "combine sales fields for readability"
- "show sales with concatenated fields"
- **What it does:** Creates readable summary fields by combining related data
#### 4. Null Value Removal/Handling
- **Query Examples:**
- "clean sales data with null handling"
- "remove nulls from sales data"
- "handle null values in sales"
- **What it does:** Filters out incomplete records and handles null values
### Regular Operations
- **"list all sales"** - Shows regular unformatted sales data
- **"show customers"** - Shows customer data
- **"list products"** - Shows product inventory
### Smart Name-Based Operations (NEW!)
- **"delete customer Alice"** - Finds and deletes Alice by name
- **"delete Alice Johnson"** - Finds customer by full name
- **"remove Johnson"** - Finds customer by last name
- **"delete product Widget"** - Finds and deletes Widget by name
- **"update price of Gadget to 25"** - Updates Gadget price to $25
- **"change email of Bob to bob@new.com"** - Updates Bob's email
""")
