👉 Try the demo: https://github.com/LM-Kit/lm-kit-net-samples/tree/main/console_net/agents/web_search_assistant

Web Search Assistant for C# .NET Applications

🎯 Purpose of the Demo

The Web Search Assistant demo shows how to use LM-Kit.NET to build a general-purpose assistant that autonomously decides when to search the web. Unlike research agents that always search, this assistant behaves like ChatGPT or Copilot: it answers casual questions from its own knowledge and only searches the web when it needs fresh, factual, or real-time information.

This is the reference demo for integrating the built-in WebSearchTool into a conversational assistant.

The sample shows how to:

Register the built-in WebSearchTool with a MultiTurnConversation in a single line of code.
Let the LLM autonomously decide when web search is needed, based on the system prompt.
Provide full call-flow visibility with color-coded streaming output (reasoning, tool calls, responses).
Swap between search providers (DuckDuckGo, Brave, Tavily, Serper) with minimal code changes.

👥 Who Should Use This Demo

Developers building chatbots: add web-grounded answers to any conversational AI application.
Product teams evaluating LM-Kit.NET: see the complete tool-calling flow from registration to autonomous invocation.
AI engineers: understand how a local LLM decides when to call external tools vs. answering from its own knowledge.
Startups and enterprises: build privacy-first assistants that run locally but can still access live web data.

🚀 What Problem It Solves

LLMs have a knowledge cutoff and can hallucinate when asked about recent events or specific facts. By equipping an assistant with web search, it can:

Answer questions about current events, news, and real-time data.
Provide accurate factual information with source citations.
Maintain the conversational fluency of a general-purpose assistant.
Avoid unnecessary searches for casual questions, saving latency and resources.

💻 Demo Application Overview

Console application that:

Lets you select a language model from a curated list (or paste a custom URI).
Creates a MultiTurnConversation with WebSearchTool and DateTimeTool registered.
Enters an interactive chat loop with full streaming output.
Color-codes every phase of the call flow so you can see exactly what happens.

✨ Key Features

One-line tool registration: chat.Tools.Register(BuiltInTools.WebSearch) is all it takes.
Autonomous tool calling: the model decides when to search based on the system prompt.
DuckDuckGo by default: no API key required, works out of the box.
Provider switching: swap to Brave, Tavily, or Serper with one line of code.
DateTimeTool: the model knows the current date for time-sensitive queries.
Call-flow visibility: reasoning (blue), tool calls (red), responses (white), stats (gray).
Special commands: /reset clears history, /regenerate retries, /continue extends the last response.

🏗️ Architecture

┌─────────────────┐
│   User Input    │
└────────┬────────┘
         │
┌────────▼────────┐
│  Web Search     │
│  Assistant      │
│  (MultiTurn     │
│   Conversation) │
└────────┬────────┘
         │
    ┌────┴────┐
    │  Tools  │
    ├─────────┤
    │WebSearch│──── DuckDuckGo / Brave / Tavily / Serper
    ├─────────┤
    │DateTime │──── Current date/time awareness
    └─────────┘

Tool Registration

using LMKit.Agents.Tools.BuiltIn;

MultiTurnConversation chat = new(model);

// One line to register web search (DuckDuckGo, no API key)
chat.Tools.Register(BuiltInTools.WebSearch);

// One line for date/time awareness
chat.Tools.Register(BuiltInTools.DateTimeNow);

Switching Providers

using LMKit.Agents.Tools.BuiltIn.Net;

// Brave Search (free tier at https://brave.com/search/api/)
var webSearch = BuiltInTools.CreateWebSearch(WebSearchTool.Provider.Brave, "YOUR_API_KEY");

// Tavily (AI-optimized search, https://tavily.com)
var webSearch = BuiltInTools.CreateWebSearch(WebSearchTool.Provider.Tavily, "YOUR_API_KEY");

// Serper (Google results via API, https://serper.dev)
var webSearch = BuiltInTools.CreateWebSearch(WebSearchTool.Provider.Serper, "YOUR_API_KEY");

⚙️ Getting Started

Prerequisites

.NET 8.0 or later
Sufficient VRAM for the selected model (4-18 GB)
Internet connectivity for web search

Download

git clone https://github.com/LM-Kit/lm-kit-net-samples
cd lm-kit-net-samples/console_net/agents/web_search_assistant

Run

dotnet build
dotnet run

Then:

Select a model by typing 0-5, or paste a custom model URI.
Start chatting. The assistant will automatically search the web when needed.
Use /reset to clear history, /regenerate to retry, or /continue to extend.

🔧 Troubleshooting

No search results: DuckDuckGo may throttle frequent requests. Space out queries or switch to Brave/Tavily.
Model does not search: try a larger model (GPT OSS 20B or GLM 4.7 Flash recommended). Smaller models may not reliably trigger tool calls.
Slow responses: web search adds network latency. Responses without search are faster.
Timeout: the default timeout is 3 minutes. For complex queries requiring multiple searches, this may not be enough.

🚀 Extend the Demo

Premium search: switch to Brave, Tavily, or Serper for higher quality results.
More tools: register additional built-in tools like BuiltInTools.CalcArithmetic or custom tools to expand capabilities.
RAG integration: combine web search with local document retrieval for hybrid grounding.
Agent framework: migrate from MultiTurnConversation to the Agent builder with PlanningStrategy.ReAct for multi-step research tasks.
Persistent sessions: save and restore conversation history across application restarts.

Table of Contents