At Konfuzio, we've seen how AI is changing document workflows. But what happens when AI even changes the way users find what they are looking for on a homepage?
"Show me the company's current ISO 27001 certificate."
This is exactly what NLWeb (Natural Language Web) from Microsoft does. Instead of clicking on drop-down menus or switching filters, users can now enter (or speak) natural questions, such as:
NLWeb from Microsoft offers an open source framework for conversational interfaces of RSS feeds from homepages, based on the MCP protocol from Anthropic.
What is NLWeb?
Microsoft has developed NLWeb an open source technology which makes it possible to use natural language directly as an interface on websites.
Instead of clicking through menus or setting filters by hand, users can now simply speak or write what they are looking for - for example:
"Show me elegant shoes under 100 euros for a summer wedding"
This not only changes the user experience, but also the way developers plan and implement applications.
Architecture of NLWeb
What's exciting about NLWeb is not just the conversational layer, but also the underlying technical architecture, which is optimized for developers who want to create smarter, AI-native web applications. Each question is evaluated by various prompts, such as the Code on GitHub shows.
More than 50 comments in The Verge interview by David Pierce with Ramanathan V. Guha from Microsoft provide an initial overview of the mood: Microsoft's plan to fix the web.
Opinions on the functionality of NLWeb are mixed: positive voices emphasize the potential benefits, such as the ability to find complex information quickly using natural language and prepare it in a structured way. In comparison, one user sees an advantage over traditional approaches, such as those used by specialized software tools.
But there are also critical voices. Critics complain, for example, that the technology in its current form pushes existing resources, such as traditional websites, into the background and thus negatively influences the visibility of individual content. Others express doubts as to whether NLWeb is able to deliver accurate and relevant results, or whether it simply replicates existing search functions in a variety of ways, but without any major added value. In particular, the difference to previously known solutions, such as extensive specialized search engines, is questioned.
NLWeb now offers website operators greater influence over the content displayed. In another article, I therefore deduce that AI a kind of target group of marketing and PR.
A technical overview
NLWeb is not just a front-end gimmick, but a server-side framework that interprets natural language, retrieves relevant information from data sources and outputs it in understandable language.
By loading the video, you accept YouTube's privacy policy.
Read more
The technological core consists of a Python server, a REST API and interfaces to Large Language Models (LLMs) such as OpenAI, Azure OpenAI or Gemini.
Central technical components:
- Configuration via YAML files (e.g. config_llm.yaml)
- Processing structured content (Schema.org, RSS, CSV)
- Indexing using vector databases (e.g. Qdrant, Milvus, Azure Search)
- Conversational retrieval with RAG technology (Retrieval Augmented Generation)
- Integration of a Model Context Protocol (MCP) that is suitable for both humans and AI agents
The inquiry process: from text to answer
With NLWeb, the path from user input to response takes place in several steps:
- The request reaches the server via an API endpoint such as /ask or /mcp/ask.
- The server loads the configurations that define which LLM and which data sources are used.
- The website content is read in, converted into vectors and stored in a database.
- The user question is analyzed, enriched with conversation history if necessary and semantically classified.
- Suitable content is retrieved from the vector database using a similarity search (Top-K).
- The results are evaluated and prioritized.
- The LLM generates a comprehensible answer based on the content.
- The response is returned to the front end in real time, and can also be streamed token by token.
Integration on the client side
Use on the website can be minimally invasive: Using a few additional HTML attributes or simple JavaScript functions, a developer can integrate a conversational interface.
Example:
curl -X POST https://yoursite.com/wp-json/nlweb/v1/ask \
-H "Content-Type: application/json" \
-d '{
"question": "What is this website about?",
"context": {
"post_type": ["post", "page"],
"category": "tutorials",
"limit": 10,
"meta_query": {
"featured": "yes"
}
}
}'Answer
{
"@context": "https://schema.org",
"@type": "SearchResultsPage",
"query": "What is this website about?",
"totalResults": 3,
"processingTime": "0.245s",
"items": [
{
"@type": "Article",
"@id": "https://yoursite.com/about/",
"name": "About Us",
"description": "Learn about our company mission and values...",
"url": "https://yoursite.com/about/",
"datePublished": "2024-01-15T10:30:00Z",
"dateModified": "2024-01-20T14:15:00Z",
"author": {
"@type": "Person",
"name": "John Doe"
},
"keywords": ["about", "company", "mission"],
"relevanceScore": 0.95
}
]
}Why this is important - for developers and UX designers
NLWeb is fundamentally changing the frontend and UX paradigm. Instead of relying on complex filters, navigation bars or category systems, the semantic content of the content and its machine-readable structure are coming to the fore.
For developers, this means
- Less logic in the front end, more importance for structured content and APIs
- Success depends on clean markup with Schema.org and semantic HTML
- Requests to the backend are no longer based on predefined parameters, but on flexible, natural statements
From the front end to the voice-controlled API
While traditional web development relies on components such as buttons, dropdowns or tabs, NLWeb thinks in terms of entities: products, recipes, real estate ads - these are all data structures that NLWeb understands and can respond to. Communication takes place via the ask API, which accesses these entities and generates suitable results.
What developers can configure
NLWeb is highly customizable: so-called prompt chains in XML files (e.g. site_type.xml) can be used to define your own intents, such as how a system should respond to unclear questions or which follow-up questions it should ask. Fallbacks can also be defined in the event of comprehension problems.
The client-side interaction can even be declarative - via attributes such as data-nlweb-intent="findProduct" UI elements can be connected to the NLWeb backend without extensive JavaScript logic.
Why NLWeb could fundamentally change the market for chat software
This could lead to a major change in the market for all chat software providers - for several profound reasons.
Firstly, NLWeb shifts the focus from specialized, mostly proprietary chat solutions to an open, system-based infrastructure that is anchored directly on the website itself. While conventional chatbots are often integrated as external widgets that rely on predefined rules or scripts, NLWeb integrates natural language directly into the semantic structure of the site. This means that the chat is no longer an additional function, but becomes the primary interface.
Secondly, user expectations are also changing. Anyone who experiences that a website understands what is meant in natural language - without complicated menu navigation or endless click paths - will react impatiently or even frustrated with classic chat solutions. This increases the pressure on existing providers to fundamentally rethink their systems: away from pure dialog guidance and towards semantic content development.
Thirdly, the opening up of Microsoft's technology as an open source project is leading to enormous pressure to innovate in the market. Previous providers of chat software, who rely heavily on closed systems or their own proprietary LLMs, are now under pressure to innovate. NLWeb significantly lowers the barriers to entry for developers - instead of buying an entire chat solution, companies can develop their own, domain-specific conversation interfaces with complete control over data, prompt design and hosting.
Fourthly, the economic model is also changing: where previously it was often billed per chat or per response (e.g. via API call to an LLM), NLWeb allows much more efficient operation thanks to targeted retrieval logic and only selective generation by the LLM. This reduces operating costs and makes it particularly attractive for medium-sized companies to become independent of traditional chat providers.
To summarize:
- NLWeb decouples chat functionality from centralized bot providers
- The interface becomes part of the website, not an add-on
- Open technology leads to lower costs and greater adaptability
- Providers of classic chat software need to reposition themselves - in terms of content, technology and price
The result: a market shakeout is likely. Providers who only see their USP in the provision of a conventional chat component will lose relevance. The winners will be those who can offer added value through data integration, semantic web understanding and domain-specific prompt engineering.
More information
The complete documentation and the code are publicly available on GitHub:
And the Microsoft feature for the launch:
The first WordPress add-in is already available via GitHub: Show HN: WPNLWeb - open-source WordPress plugin using Microsoft's NLWeb Protocol | Hacker News
EN 
DE 
ES 
FR