Llm

See also [LLM 0.13: The annotated release notes](https://simonwillison.net/2024/Jan/26/llm/).

- Added support for new OpenAI embedding models: `3-small` and `3-large` and three variants of those with different dimension sizes,
`3-small-512`, `3-large-256` and `3-large-1024`. See {ref}`OpenAI embedding models <openai-models-embedding>` for details. [394](https://github.com/simonw/llm/issues/394)
- The default `gpt-4-turbo` model alias now points to `gpt-4-turbo-preview`, which uses the most recent OpenAI GPT-4 turbo model (currently `gpt-4-0125-preview`). [396](https://github.com/simonw/llm/issues/396)
- New OpenAI model aliases `gpt-4-1106-preview` and `gpt-4-0125-preview`.
- OpenAI models now support a `-o json_object 1` option which will cause their output to be returned as a valid JSON object. [373](https://github.com/simonw/llm/issues/373)
- New {ref}`plugins <plugin-directory>` since the last release include [llm-mistral](https://github.com/simonw/llm-mistral), [llm-gemini](https://github.com/simonw/llm-gemini), [llm-ollama](https://github.com/taketwo/llm-ollama) and [llm-bedrock-meta](https://github.com/flabat/llm-bedrock-meta).
- The `keys.json` file for storing API keys is now created with `600` file permissions. [351](https://github.com/simonw/llm/issues/351)
- Documented {ref}`a pattern <homebrew-warning>` for installing plugins that depend on PyTorch using the Homebrew version of LLM, despite Homebrew using Python 3.12 when PyTorch have not yet released a stable package for that Python version. [397](https://github.com/simonw/llm/issues/397)
- Underlying OpenAI Python library has been upgraded to `>1.0`. It is possible this could cause compatibility issues with LLM plugins that also depend on that library. [325](https://github.com/simonw/llm/issues/325)
- Arrow keys now work inside the `llm chat` command. [376](https://github.com/simonw/llm/issues/376)
- `LLM_OPENAI_SHOW_RESPONSES=1` environment variable now outputs much more detailed information about the HTTP request and response made to OpenAI (and OpenAI-compatible) APIs. [404](https://github.com/simonw/llm/issues/404)
- Dropped support for Python 3.7.

(v0_12)=

- Support for the [new GPT-4 Turbo model](https://openai.com/blog/new-models-and-developer-products-announced-at-devday) from OpenAI. Try it using `llm chat -m gpt-4-turbo` or `llm chat -m 4t`. [#323](https://github.com/simonw/llm/issues/323)
- New `-o seed 1` option for OpenAI models which sets a seed that can attempt to evaluate the prompt deterministically. [324](https://github.com/simonw/llm/issues/324)

(v0_11_2)=

- Pin to version of OpenAI Python library prior to 1.0 to avoid breaking. [327](https://github.com/simonw/llm/issues/327)

(v0_11_1)=

- Fixed a bug where `llm embed -c "text"` did not correctly pick up the configured {ref}`default embedding model <embeddings-cli-embed-models-default>`. [317](https://github.com/simonw/llm/issues/317)
- New plugins: [llm-python](https://github.com/simonw/llm-python), [llm-bedrock-anthropic](https://github.com/sblakey/llm-bedrock-anthropic) and [llm-embed-jina](https://github.com/simonw/llm-embed-jina) (described in [Execute Jina embeddings with a CLI using llm-embed-jina](https://simonwillison.net/2023/Oct/26/llm-embed-jina/)).
- [llm-gpt4all](https://github.com/simonw/llm-gpt4all) now uses the new GGUF model format. [simonw/llm-gpt4all#16](https://github.com/simonw/llm-gpt4all/issues/16)

(v0_11)=

LLM now supports the new OpenAI `gpt-3.5-turbo-instruct` model, and OpenAI completion (as opposed to chat completion) models in general. [284](https://github.com/simonw/llm/issues/284)

bash
llm -m gpt-3.5-turbo-instruct 'Reasons to tame a wild beaver:'

OpenAI completion models like this support a `-o logprobs 3` option, which accepts a number between 1 and 5 and will include the log probabilities (for each produced token, what were the top 3 options considered by the model) in the logged response.

bash
llm -m gpt-3.5-turbo-instruct 'Say hello succinctly' -o logprobs 3

You can then view the `logprobs` that were recorded in the SQLite logs database like this:
bash
sqlite-utils "$(llm logs path)" \
'select * from responses order by id desc limit 1' | \
jq '.[0].response_json' -r | jq

Truncated output looks like this:

[
{
"text": "Hi",
"top_logprobs": [
{
"Hi": -0.13706253,
"Hello": -2.3714375,
"Hey": -3.3714373
}
]
},
{
"text": " there",
"top_logprobs": [
{
" there": -0.96057636,
"!\"": -0.5855763,
".\"": -3.2574513
}
]
}
]

Also in this release:

- The `llm.user_dir()` function, used by plugins, now ensures the directory exists before returning it. [275](https://github.com/simonw/llm/issues/275)
- New `LLM_OPENAI_SHOW_RESPONSES=1` environment variable for displaying the full HTTP response returned by OpenAI compatible APIs. [286](https://github.com/simonw/llm/issues/286)
- The `llm embed-multi` command now has a `--batch-size X` option for setting the batch size to use when processing embeddings - useful if you have limited memory available. [273](https://github.com/simonw/llm/issues/273)
- The `collection.embed_multi()` method also now accepts an optional `batch_size=int` argument.
- Fixed two bugs with `llm embed-multi --files` relating to handling of directories. Thanks, [ealvar3z](https://github.com/ealvar3z). [#274](https://github.com/simonw/llm/issues/274), [#280](https://github.com/simonw/llm/issues/280)

(v0_10)=

The two major features in this release are the **{ref}`llm chat command <usage-chat>`** and support for **embedding binary data**.

See [Build an image search engine with llm-clip, chat with models with llm chat](https://simonwillison.net/2023/Sep/12/llm-clip-and-chat/) for more background on these features.

llm chat

The new `llm chat` command starts an ongoing chat conversation with a model in your terminal. It works with all models supported by LLM and its {ref}`plugins <plugins>`, including locally installed models such as Llama 2. [231](https://github.com/simonw/llm/issues/231)

This offers a big performance boost for local models, since they don't need to be freshly loaded into memory for each prompt.

Here's an example chat with Llama 2 13B, provided by the [llm-mlc](https://github.com/simonw/llm-mlc) plugin.

bash
llm chat -m mlc-chat-Llama-2-13b-chat-hf-q4f16_1

This starts a chat session:

Type 'exit' or 'quit' to exit
Type '!multi' to enter multiple lines, then '!end' to finish
> Who are you?
Hello! I'm just an AI, here to assist you with any questions you may have.
My name is LLaMA, and I'm a large language model trained to provide helpful
and informative responses to a wide range of questions and topics. I'm here
to help you find the information you need, while ensuring a safe and
respectful conversation. Please feel free to ask me anything, and I'll do my
best to provide a helpful and accurate response.
> Tell me a joke about otters
Sure, here's a joke about otters:

Why did the otter go to the party?

Because he heard it was a "whale" of a time!

(Get it? Whale, like a big sea mammal, but also a "wild" or "fun" time.
Otters are known for their playful and social nature, so it's a lighthearted
and silly joke.)

I hope that brought a smile to your face! Do you have any other questions or
topics you'd like to discuss?
> exit

Chat sessions are {ref}`logged to SQLite <logging>` - use `llm logs` to view them. They can accept system prompts, templates and model options - consult {ref}`the chat documentation <usage-chat>` for details.

Binary embedding support

LLM's {ref}`embeddings feature <embeddings>` has been expanded to provide support for embedding binary data, in addition to text. [254](https://github.com/simonw/llm/pull/254)

This enables models like [CLIP](https://openai.com/research/clip), supported by the new **[llm-clip](https://github.com/simonw/llm-clip)** plugin.

CLIP is a multi-modal embedding model which can embed images and text into the same vector space. This means you can use it to create an embedding index of photos, and then search for the embedding vector for "a happy dog" and get back images that are semantically closest to that string.

To create embeddings for every JPEG in a directory stored in a `photos` collection, run:

bash
llm install llm-clip
llm embed-multi photos --files photos/ '*.jpg' --binary -m clip

Now you can search for photos of raccoons using:

llm similar photos -c 'raccoon'

This spits out a list of images, ranked by how similar they are to the string "raccoon":

{"id": "IMG_4801.jpeg", "score": 0.28125139257127457, "content": null, "metadata": null}
{"id": "IMG_4656.jpeg", "score": 0.26626441704164294, "content": null, "metadata": null}
{"id": "IMG_2944.jpeg", "score": 0.2647445926996852, "content": null, "metadata": null}
...


Also in this release

- The {ref}`LLM_LOAD_PLUGINS environment variable <llm-load-plugins>` can be used to control which plugins are loaded when `llm` starts running. [256](https://github.com/simonw/llm/issues/256)
- The `llm plugins --all` option includes builtin plugins in the list of plugins. [259](https://github.com/simonw/llm/issues/259)
- The `llm embed-db` family of commands has been renamed to `llm collections`. [229](https://github.com/simonw/llm/issues/229)
- `llm embed-multi --files` now has an `--encoding` option and defaults to falling back to `latin-1` if a file cannot be processed as `utf-8`. [225](https://github.com/simonw/llm/issues/225)

(v0_10_a1)=