Powering Natural Language Interfaces for the Web with NLWeb and Snowflake

Today we’re excited to announce Snowflake’s integration with NLWeb, an open-source project from Microsoft that enables developers to build natural language interfaces over website content using structured data and modular, pluggable AI components. 

As a partner, Snowflake allows developers to use several of its core AI infrastructure products to power NLWeb: Cortex LLM REST APIs for embedding and chat inference, as well as Cortex Search for fast high-quality retrieval over unstructured data.

Through this integration, web developers can easily spin up a conversational AI application on their website data in minutes, using Snowflake retrieval and inference infrastructure for chatbot context retrieval and response generation. For Snowflake customers, it is now seamless to power conversational web interfaces using secure, scalable AI and data infrastructure.

NLWeb is a project that provides a framework that allows developers to easily turn websites into AI applications. It is published in a public Github repository.

At a high level, NLWeb works by:

• Parsing and ingesting data in existing semi-structured formats like Schema.org, RSS, and other data that websites already publish

• Indexing that content in a vector retrieval service

• Routing user queries through a lightweight server that handles natural language understanding and retrieval

• Optionally augmenting or answering queries with an LLM

The result is a high-quality natural language interface on top of web data, giving developers the ability to let users “talk to” web data. 

By default, every NLWeb instance is MCP compatible, meaning that once you stand up a chatbot on your website, you can easily expose that chat interface to other MCP-compatible agentic applications. Additionally, NLWeb is modular by design, so developers can bring their own retrieval system, model APIs, and define their own extensions.

The basic value proposition of NLWeb is to very easily empower developers with existing websites or APIs to create chatbots over their web data. If you’re a website owner and you want to let users “chat” with your data – e.g., for eCommerce product discovery or simplify news content consumption – NLWeb gives you a simple recipe for doing so.

Traditionally, web content has historically been optimized for human navigation via menus and search bars. However, LLMs allow us to interact with web content in a chat interface. With NLWeb, you can now easily turn websites into composable natural language endpoints that are interoperable with both users and agents. NLWeb provides a way for developers to make their websites more discoverable, accessible, and personalized for their end-consumers.

Snowflake is integrated into the NLWeb reference stack in two places, making Snowflake a one-stop-shop for supporting NLWeb instances with scalable AI infrastructure:

1. Retrieval with Cortex Search

Cortex Search is Snowflake’s fully-managed hybrid search service for large-scale, low-latency natural language retrieval. It uses a hybrid architecture that combines semantic vector search, keyword match, and semantic reranking to return highly relevant documents from unstructured and semi-structured corpora. 

Through this collaboration, users with web data stored in Snowflake can build NLWeb instances directly on top of that Snowflake data in Cortex Search instead of relying on external crawlers to populate the vector store. This makes it easy and secure for Snowflake developers to power NLWeb instances on top of their existing web data.

2. LLM Inference with Cortex LLM APIs

Cortex LLM APIs provide hosted access to serverless, high-performance language model inference functions directly within Snowflake. Cortex hosts generative models from popular providers like Anthropic, OpenAI, Meta’s Llama, Mistral, DeepSeek, and more. In addition, Cortex provides hosted embedding functions for Snowflake’s Arctic models, Voyage models, and other popular open-source models.

With this integration, developers of NLWeb instances can power chatbot inference with Cortex LLM APIs in a secure and scalable fashion. No additional hosting or orchestration is required.

Snowflake is supportive of NLWeb’s mission of bringing delightful chat experiences to every website and content publisher. To that end, Snowflake customers can easily spin up NLWeb instances on data that is secure and governed within Snowflake without worrying about retrieval and LLM inference infrastructure.

The NLWeb GitHub repo includes:

• A reference server for handling natural language queries

• Integration with Cortex Search and Cortex LLM APIs

  • Cortex Search integration

  • Cortex LLM inference API integration

  • Cortex embedding inference API integration