This library provides an interface for interacting with Large Language Models (LLMs). It allows elisp code to use LLMs while also giving end-users the choice to select their preferred LLM. This is particularly beneficial when working with LLMs since various high-quality models exist, some of which have paid API access, while others are locally installed and free but offer medium quality. Applications using LLMs can utilize this library to ensure compatibility regardless of whether the user has a local LLM or is paying for API access.

LLMs exhibit varying functionalities and APIs. This library aims to abstract functionality to a higher level, as some high-level concepts might be supported by an API while others require more low-level implementations. An example of such a concept is “examples,” where the client offers example interactions to demonstrate a pattern for the LLM. While the GCloud Vertex API has an explicit API for examples, OpenAI’s API requires specifying examples by modifying the system prompt. OpenAI also introduces the concept of a system prompt, which does not exist in the Vertex API. Our library aims to conceal these API variations by providing higher-level concepts in our API.

Certain functionalities might not be available in some LLMs. Any such unsupported functionality will raise a 'not-implemented signal.

Users of an application that uses this package should not need to install it themselves. The llm package should be installed as a dependency when you install the package that uses it. However, you do need to require the llm module and set up the provider you will be using. Typically, applications will have a variable you can set. For example, let’s say there’s a package called “llm-refactoring”, which has a variable llm-refactoring-provider. You would set it up like so:

(use-package llm-refactoring
  :init
  (require 'llm-openai)
  (setq llm-refactoring-provider (make-llm-openai :key my-openai-key))

Here my-openai-key would be a variable you set up before with your OpenAI key. Or, just substitute the key itself as a string. It’s important to remember never to check your key into a public repository such as GitHub, because your key must be kept private. Anyone with your key can use the API, and you will be charged.

All of the providers (except for llm-fake), can also take default parameters that will be used if they are not specified in the prompt. These are the same parameters as appear in the prompt, but prefixed with default-chat-. So, for example, if you find that you like Ollama to be less creative than the default, you can create your provider like:

(make-llm-ollama :embedding-model "mistral:latest" :chat-model "mistral:latest" :default-chat-temperature 0.1)

For embedding users. if you store the embeddings, you must set the embedding model. Even though there’s no way for the llm package to tell whether you are storing it, if the default model changes, you may find yourself storing incompatible embeddings.

You can set up with make-llm-openai, with the following parameters:

• :key, the Open AI key that you get when you sign up to use Open AI’s APIs. Remember to keep this private. This is non-optional.
• :chat-model: A model name from the list of Open AI’s model names. Keep in mind some of these are not available to everyone. This is optional, and will default to a reasonable 3.5 model.
• :embedding-model: A model name from list of Open AI’s embedding model names. This is optional, and will default to a reasonable model.

There are many Open AI compatible APIs and proxies of Open AI. You can set up one with make-llm-openai-compatible, with the following parameter:

• :url, the URL of leading up to the command (“embeddings” or “chat/completions”). So, for example, “https://api.openai.com/v1/” is the URL to use Open AI (although if you wanted to do that, just use make-llm-openai instead.

Microsoft Azure has an Open AI integration, although it doesn’t support everything Open AI does, such as function calling. You can set it up with make-llm-azure, with the following parameter:

• :url, the endpoint URL, such as “https://docs-test-001.openai.azure.com/”.

This is Google’s AI model. You can get an API key via their page on Google AI Studio. Set this up with make-llm-gemini, with the following parameters:

• :key, the Google AI key that you get from Google AI Studio.
• :chat-model, the model name, from the [[https://ai.google.dev/models][list of models. This is optional and will default to the text Gemini model.
• :embedding-model: the model name, currently must be “embedding-001”. This is optional and will default to “embedding-001”.

This is mostly for those who want to use Google Cloud specifically, most users should use Gemini instead, which is easier to set up.

You can set up with make-llm-vertex, with the following parameters:

• :project: Your project number from Google Cloud that has Vertex API enabled.
• :chat-model: A model name from the list of Vertex’s model names. This is optional, and will default to a reasonable model.
• :embedding-model: A model name from the list of Vertex’s embedding model names. This is optional, and will default to a reasonable model.

In addition to the provider, which you may want multiple of (for example, to charge against different projects), there are customizable variables:

• llm-vertex-gcloud-binary: The binary to use for generating the API key.
• llm-vertex-gcloud-region: The gcloud region to use. It’s good to set this to a region near where you are for best latency. Defaults to “us-central1”.

  If you haven’t already, you must run the following command before using this:

  gcloud beta services identity create --service=aiplatform.googleapis.com --project=PROJECT_ID
      

Claude is Anthropic’s large language model. It does not support embeddings. It does support function calling, but currently not in streaming. You can set it up with the following parameters:

:key: The API key you get from Claude’s settings page. This is required. :chat-model: One of the Claude models. Defaults to “claude-3-opus-20240229”, the most powerful model.

Ollama is a way to run large language models locally. There are many different models you can use with it, and some of them support function calling. You set it up with the following parameters:

• :scheme: The scheme (http/https) for the connection to ollama. This default to “http”.
• :host: The host that ollama is run on. This is optional and will default to localhost.
• :port: The port that ollama is run on. This is optional and will default to the default ollama port.
• :chat-model: The model name to use for chat. This is not optional for chat use, since there is no default.
• :embedding-model: The model name to use for embeddings (only [some models](https://ollama.com/search?q=&c=embedding) can be used for embeddings. This is not optional for embedding use, since there is no default.

GPT4All is a way to run large language models locally. To use it with llm package, you must click “Enable API Server” in the settings. It does not offer embeddings or streaming functionality, though, so Ollama might be a better fit for users who are not already set up with local models. You can set it up with the following parameters:

• :host: The host that GPT4All is run on. This is optional and will default to localhost.
• :port: The port that GPT4All is run on. This is optional and will default to the default ollama port.
• :chat-model: The model name to use for chat. This is not optional for chat use, since there is no default.

llama.cpp is a way to run large language models locally. To use it with the llm package, you need to start the server (with the “–embedding” flag if you plan on using embeddings). The server must be started with a model, so it is not possible to switch models until the server is restarted to use the new model. As such, model is not a parameter to the provider, since the model choice is already set once the server starts.

There is a deprecated provider, however it is no longer needed. Instead, llama cpp is Open AI compatible, so the Open AI Compatible provider should work.

This is a client that makes no call, but it just there for testing and debugging. Mostly this is of use to programmatic clients of the llm package, but end users can also use it to understand what will be sent to the LLMs. It has the following parameters:

• :output-to-buffer: if non-nil, the buffer or buffer name to append the request sent to the LLM to.
• :chat-action-func: a function that will be called to provide a string or symbol and message cons which are used to raise an error.
• :embedding-action-func: a function that will be called to provide a vector or symbol and message cons which are used to raise an error.

When picking a chat or embedding model, anything can be used, as long as the service thinks it is valid. However, models vary on context size and capabilities. The llm-prompt module, and any client, can depend on the context size of the model via llm-chat-token-limit. Similarly, some models have different capabilities, exposed in llm-capabilities. The llm-models module defines a list of popular models, but this isn’t a comprehensive list. If you want to add a model, it is fairly easy to do, for example here is adding the Mistral model (which is already included, though):

(require 'llm-models)
(add-to-list
 'llm-models
 (make-llm-model
  :name "Mistral" :symbol 'mistral
  :capabilities '(generation tool-use free-software)
  :context-length 8192
  :regex "mistral"))

The :regex needs to uniquely identify the model passed in from a provider’s chat or embedding model.

Once this is done, the model will be recognized to have the given context length and capabilities.

The llm package is part of GNU Emacs by being part of GNU ELPA. Unfortunately, the most popular LLMs in use are non-free, which is not what GNU software should be promoting by inclusion. On the other hand, by use of the llm package, the user can make sure that any client that codes against it will work with free models that come along. It’s likely that sophisticated free LLMs will, emerge, although it’s unclear right now what free software means with respsect to LLMs. Because of this tradeoff, we have decided to warn the user when using non-free LLMs (which is every LLM supported right now except the fake one). You can turn this off the same way you turn off any other warning, by clicking on the left arrow next to the warning when it comes up. Alternatively, you can set llm-warn-on-nonfree to nil. This can be set via customization as well.

To build upon the example from before:

(use-package llm-refactoring
  :init
  (require 'llm-openai)
  (setq llm-refactoring-provider (make-llm-openai :key my-openai-key)
        llm-warn-on-nonfree nil)

Client applications should require the llm package, and code against it. Most functions are generic, and take a struct representing a provider as the first argument. The client code, or the user themselves can then require the specific module, such as llm-openai, and create a provider with a function such as (make-llm-openai :key user-api-key). The client application will use this provider to call all the generic functions.

For all callbacks, the callback will be executed in the buffer the function was first called from. If the buffer has been killed, it will be executed in a temporary buffer instead.

• llm-chat provider prompt: With user-chosen provider , and a llm-chat-prompt structure (created by llm-make-chat-prompt), send that prompt to the LLM and wait for the string output.
• llm-chat-async provider prompt response-callback error-callback: Same as llm-chat, but executes in the background. Takes a response-callback which will be called with the text response. The error-callback will be called in case of error, with the error symbol and an error message.
• llm-chat-streaming provider prompt partial-callback response-callback error-callback: Similar to llm-chat-async, but request a streaming response. As the response is built up, partial-callback is called with the all the text retrieved up to the current point. Finally, reponse-callback is called with the complete text.
• llm-embedding provider string: With the user-chosen provider, send a string and get an embedding, which is a large vector of floating point values. The embedding represents the semantic meaning of the string, and the vector can be compared against other vectors, where smaller distances between the vectors represent greater semantic similarity.
• llm-embedding-async provider string vector-callback error-callback: Same as llm-embedding but this is processed asynchronously. vector-callback is called with the vector embedding, and, in case of error, error-callback is called with the same arguments as in llm-chat-async.
• llm-batch-embedding provider strings: same as llm-embedding, but takes in a list of strings, and returns a list of vectors whose order corresponds to the ordering of the strings.
• llm-batch-embedding-async provider strings vectors-callback error-callback: same as llm-embedding-async, but takes in a list of strings, and returns a list of vectors whose order corresponds to the ordering of the strings.
• llm-count-tokens provider string: Count how many tokens are in string. This may vary by provider, because some provideres implement an API for this, but typically is always about the same. This gives an estimate if the provider has no API support.
• llm-cancel-request request Cancels the given request, if possible. The request object is the return value of async and streaming functions.
• llm-name provider. Provides a short name of the model or provider, suitable for showing to users.
• llm-chat-token-limit. Gets the token limit for the chat model. This isn’t possible for some backends like llama.cpp, in which the model isn’t selected or known by this library.

  And the following helper functions:

  • llm-make-chat-prompt text &keys context examples functions temperature max-tokens: This is how you make prompts. text can be a string (the user input to the llm chatbot), or a list representing a series of back-and-forth exchanges, of odd number, with the last element of the list representing the user’s latest input. This supports inputting context (also commonly called a system prompt, although it isn’t guaranteed to replace the actual system prompt), examples, and other important elements, all detailed in the docstring for this function. The non-standard-params let you specify other options that might vary per-provider. The correctness is up to the client.
  • llm-chat-prompt-to-text prompt: From a prompt, return a string representation. This is not usually suitable for passing to LLMs, but for debugging purposes.
  • llm-chat-streaming-to-point provider prompt buffer point finish-callback: Same basic arguments as llm-chat-streaming, but will stream to point in buffer.
  • llm-chat-prompt-append-response prompt response role: Append a new response (from the user, usually) to the prompt. The role is optional, and defaults to 'user.

Interactions with the llm package can be logged by setting llm-log to a non-nil value. This should be done only when developing. The log can be found in the *llm log* buffer.

Conversations can take place by repeatedly calling llm-chat and its variants. The prompt should be constructed with llm-make-chat-prompt. For a conversation, the entire prompt must be kept as a variable, because the llm-chat-prompt-interactions slot will be getting changed by the chat functions to store the conversation. For some providers, this will store the history directly in llm-chat-prompt-interactions, but other LLMs have an opaque conversation history. For that reason, the correct way to handle a conversation is to repeatedly call llm-chat or variants with the same prompt structure, kept in a variable, and after each time, add the new user text with llm-chat-prompt-append-response. The following is an example:

(defvar-local llm-chat-streaming-prompt nil)
(defun start-or-continue-conversation (text)
  "Called when the user has input TEXT as the next input."
  (if llm-chat-streaming-prompt
      (llm-chat-prompt-append-response llm-chat-streaming-prompt text)
    (setq llm-chat-streaming-prompt (llm-make-chat-prompt text))
    (llm-chat-streaming-to-point provider llm-chat-streaming-prompt (current-buffer) (point-max) (lambda ()))))

The interactions in a prompt may be modified by conversation or by the conversion of the context and examples to what the LLM understands. Different providers require different things from the interactions. Some can handle system prompts, some cannot. Some require alternating user and assistant chat interactions, others can handle anything. It’s important that clients keep to behaviors that work on all providers. Do not attempt to read or manipulate llm-chat-prompt-interactions after initially setting it up for the first time, because you are likely to make changes that only work for some providers. Similarly, don’t directly create a prompt with make-llm-chat-prompt, because it is easy to create something that wouldn’t work for all providers.

Note: function calling functionality is currently alpha quality. If you want to use function calling, please watch the =llm= discussions for any announcements about changes.

Function calling is a way to give the LLM a list of functions it can call, and have it call the functions for you. The standard interaction has the following steps:

1. The client sends the LLM a prompt with functions it can call.
2. The LLM may return which functions to execute, and with what arguments, or text as normal.
3. If the LLM has decided to call one or more functions, those functions should be called, and their results sent back to the LLM.
4. The LLM will return with a text response based on the initial prompt and the results of the function calling.
5. The client can now can continue the conversation.

This basic structure is useful because it can guarantee a well-structured output (if the LLM does decide to call the function). Not every LLM can handle function calling, and those that do not will ignore the functions entirely. The function llm-capabilities will return a list with function-calls in it if the LLM supports function calls. Right now only Gemini, Vertex, Claude, and Open AI support function calling. Ollama should get function calling soon. However, even for LLMs that handle function calling, there is a fair bit of difference in the capabilities. Right now, it is possible to write function calls that succeed in Open AI but cause errors in Gemini, because Gemini does not appear to handle functions that have types that contain other types. So client programs are advised for right now to keep function to simple types.

The way to call functions is to attach a list of functions to the llm-function-call slot in the prompt. This is a list of llm-function-call structs, which takes a function, a name, a description, and a list of llm-function-arg structs. The docstrings give an explanation of the format.

The various chat APIs will execute the functions defined in llm-function-call with the arguments supplied by the LLM. Instead of returning (or passing to a callback) a string, instead an alist will be returned of function names and return values.

After sending a function call, the client could use the result, but if you want to proceed with the conversation, or get a textual response that accompany the function you should just send the prompt back with no modifications. This is because the LLM gives the function call to make as a response, and then expects to get back the results of that function call. The results were already executed at the end of the previous call, which also stores the result of that execution in the prompt. This is why it should be sent back without further modifications.

Be aware that there is no gaurantee that the function will be called correctly. While the LLMs mostly get this right, they are trained on Javascript functions, so imitating Javascript names is recommended. So, “write_email” is a better name for a function than “write-email”.

Examples can be found in llm-tester. There is also a function call to generate function calls from existing elisp functions in utilities/elisp-to-function-call.el.

The llm-prompt module provides helper functions to create prompts that can incorporate data from your application. In particular, this should be very useful for application that need a lot of context.

A prompt defined with llm-prompt is a template, with placeholders that the module will fill in. Here’s an example of a prompt definition, from the ekg package:

(llm-defprompt ekg-llm-fill-prompt
  "The user has written a note, and would like you to append to it,
to make it more useful.  This is important: only output your
additions, and do not repeat anything in the user's note.  Write
as a third party adding information to a note, so do not use the
first person.

First, I'll give you information about the note, then similar
other notes that user has written, in JSON.  Finally, I'll give
you instructions.  The user's note will be your input, all the
rest, including this, is just context for it.  The notes given
are to be used as background material, which can be referenced in
your answer.

The user's note uses tags: {{tags}}.  The notes with the same
tags, listed here in reverse date order: {{tag-notes:10}}

These are similar notes in general, which may have duplicates
from the ones above: {{similar-notes:1}}

This ends the section on useful notes as a background for the
note in question.

Your instructions on what content to add to the note:

{{instructions}}
")

When this is filled, it is done in the context of a provider, which has a known context size (via llm-chat-token-limit). Care is taken to not overfill the context, which is checked as it is filled via llm-count-tokens. We usually want to not fill the whole context, but instead leave room for the chat and subsequent terms. The variable llm-prompt-default-max-pct controls how much of the context window we want to fill. The way we estimate the number of tokens used is quick but inaccurate, so limiting to less than the maximum context size is useful for guarding against a miscount leading to an error calling the LLM due to too many tokens. If you want to have a hard limit as well that doesn’t depend on the context window size, you can use llm-prompt-default-max-tokens. We will use the minimum of either value.

Variables are enclosed in double curly braces, like this: {{instructions}}. They can just be the variable, or they can also denote a number of tickets, like so: {{tag-notes:10}}. Tickets should be thought of like lottery tickets, where the prize is a single round of context filling for the variable. So the variable tag-notes gets 10 tickets for a drawing. Anything else where tickets are unspecified (unless it is just a single variable, which will be explained below) will get a number of tickets equal to the total number of specified tickets. So if you have two variables, one with 1 ticket, one with 10 tickets, one will be filled 10 times more than the other. If you have two variables, one with 1 ticket, one unspecified, the unspecified one will get 1 ticket, so each will have an even change to get filled. If no variable has tickets specified, each will get an equal chance. If you have one variable, it could have any number of tickets, but the result would be the same, since it would win every round. This algorithm is the contribution of David Petrou.

The above is true of variables that are to be filled with a sequence of possible values. A lot of LLM context filling is like this. In the above example, {{similar-notes}} is a retrieval based on a similarity score. It will continue to fill items from most similar to least similar, which is going to return almost everything the ekg app stores. We want to retrieve only as needed. Because of this, the llm-prompt module takes in generators to supply each variable. However, a plain list is also acceptable, as is a single value. Any single value will not enter into the ticket system, but rather be prefilled before any tickets are used.

Values supplied in either the list or generators can be the values themselves, or conses. If a cons, the variable to fill is the car of the cons, and the cdr is the place to fill the new value, front or back. The front is the default: new values will be appended to the end. back will add new values to the start of the filled text for the variable instead.

So, to illustrate with this example, here’s how the prompt will be filled:

1. First, the {{tags}} and {{instructions}} will be filled first. This will happen regardless before we check the context size, so the module assumes that these will be small and not blow up the context.
2. Check the context size we want to use (llm-prompt-default-max-pct multiplied by llm-chat-token-limit) and exit if exceeded.
3. Run a lottery with all tickets and choose one of the remaining variables to fill.
4. If the variable won’t make the text too large, fill the variable with one entry retrieved from a supplied generator, otherwise ignore. These are values are not conses, so values will be appended to the end of the generated text for each variable (so a new variable generated for tags will append after other generated tags but before the subsequent “and” in the text.
5. Goto 2

The prompt can be filled two ways, one using predefined prompt template (llm-defprompt and llm-prompt-fill), the other using a prompt template that is passed in (llm-prompt-fill-text).

(llm-defprompt my-prompt "My name is {{name}} and I'm here's to say {{messages}}")

(llm-prompt-fill 'my-prompt my-llm-provider :name "Pat" :messages #'my-message-retriever)

(iter-defun my-message-retriever ()
  "Return the messages I like to say."
  (my-message-reset-messages)
  (while (my-has-next-message)
    (iter-yield (my-get-next-message))))

Alternatively, you can just fill it directly:

(llm-prompt-fill-text "Hi, I'm {{name}} and I'm here to say {{messages}}"
                      :name "John" :messages #'my-message-retriever)

As you can see in the examples, the variable values are passed in with matching keys.

If you are interested in creating a provider, please send a pull request, or open a bug. This library is part of GNU ELPA, so any major provider that we include in this module needs to be written by someone with FSF papers. However, you can always write a module and put it on a different package archive, such as MELPA.