Welcome to the new Golem Cloud Docs! 👋
Common Guides
Using LLMs

Talking to Large Language Models

Golem provides a library called golem-llm (opens in a new tab) that is a WebAssembly Component allowing Golem components written in any of the supported languages to use a set of supported LLM providers.

The current list of supported LLMs are the following:

The library contains a WIT interface (opens in a new tab) providing a unified interface for all these providers.

Starting with the LLM templates

There are example templates provided by golem component new for some languages, called example-gitpulse-agent. To start from this example from scratch, use the following sequence of golem CLI commands:

golem app new my-agent python
cd my-agent
golem component new python:example-gitpulse-agent

Adding LLM support to an existing component

To add the golem-llm dependency to an existing Golem component, follow following steps:

Select one of the released WASM binaries

Go to the releases page of golem-llm (opens in a new tab) and copy the link of to the latest WASM binary of the desired LLM provider. The -portable versions are to be used outside of Golem so choose the non-portable ones.

Add a new dependency to golem.yaml

In the golem.yaml of the component, add the following new dependency:

dependencies:
  my:component:  # <-- the name of your component you add the golem-llm dependency to
    - type: wasm
      url: https://github.com/golemcloud/golem-llm/releases/download/v0.1.2/golem_llm_openai.wasm

Import golem-llm in WIT

Add the following import statement to your component's WIT file:

import golem:llm/llm@1.0.0;

Rebuild the application

The next build will download the WASM component and regenerate the language-specific bindings for the new import:

golem app build

Using the library

There are three top level functions exported from golem-llm:

send: func(messages: list<message>, config: config) -> chat-event;
continue: func(messages: list<message>, tool-results: list<tuple<tool-call, tool-result>>, config: config) -> chat-even;
%stream: func(messages: list<message>, config: config) -> chat-stream;
  • send sends a prompt with a configuration to the LLM and returns the whole response
  • continue can be used after one send and zero or more continue calls, if the returned result contained a tool call. After performing the call, the results can be passed back to the LLM with the tool-results parameter.
  • stream returns a stream resource which can be used to process the LLM response incrementally as it arrives

For general information about how to construct the prompts and how tool usage works, read the chosen LLMs official documentation.

Images are supported in the input as messages referring to external images via an url.

Simple text chat example

The following example demonstrates the usage of send:

  use crate::bindings::golem::llm::llm;
 
  let config = llm::Config {
          model: "gpt-3.5-turbo",
          temperature: Some(0.2),
          max_tokens: None,
          stop_sequences: None,
          tools: vec![],
          tool_choice: None,
          provider_options: vec![],
  };
 
  let response = llm::send(
      &[llm::Message {
          role: llm::Role::User,
          name: Some("user-name".to_string()),
          content: vec![llm::ContentPart::Text(
              "What is the usual weather on the Vršič pass in the beginning of May?".to_string(),
          )],
      }],
      &config,
  );
Worker FilesytemUsing Relational Databases