# A Prolog library for interfacing with LLMs

> Source: <https://github.com/vagos/llmpl>
> Published: 2026-07-09 13:52:12+00:00

Use LLMs inside Prolog!

`pllm`

is a minimal SWI-Prolog helper that exposes `llm/2`

.
The predicate posts a prompt to an HTTP LLM endpoint and unifies the model's
response text with the second argument.

The library currently supports any OpenAI-compatible chat/completions endpoint.

```
?- pack_install(pllm).
```

Some services require an API key for authentication.
Set the `LLM_API_KEY`

environment variable to your API key.
You can do the following in your shell before starting SWI-Prolog:

```
echo LLM_API_KEY="sk-..." >> .env
set -a && source .env && set +a
```

Configure the endpoint and default model before calling `llm/2`

or `llm/3`

:

```
?- config("https://api.openai.com/v1/chat/completions", "gpt-4o-mini").
```

You can override the configured model per call with `llm/3`

options.

```
# Fill in .env with your settings
set -a && souce .env && set +a
swipl
?- [prolog/llm].
?- llm("Say hello in French.", Output).
Output = "Bonjour !".

?- llm("Say hello in French.", Output, [model("gpt-4o-mini"), timeout(30)]).
Output = "Bonjour !".

?- llm(Prompt, "Dog").
Prompt = "What animal is man's best friend?",
...
```

This library expects an OpenAI-compatible chat/completions endpoint. Below are common providers and endpoints you can try.

OpenAI

- Endpoint:
`https://api.openai.com/v1/chat/completions`

- Example:
`?- config("https://api.openai.com/v1/chat/completions", "gpt-4o-mini").`

Ollama (local)

- Endpoint:
`http://localhost:11434/v1/chat/completions`

- Example:
`?- config("http://localhost:11434/v1/chat/completions", "llama3.1").`

If you call `llm/2`

with an unbound first argument and a concrete response,
the library first asks the LLM to suggest a prompt that would (ideally)
produce that response, binds it to your variable, and then sends a *second*
request that wraps the suggested prompt in a hard constraint (`"answer only with ..."`

).
This costs two API calls and is still best-effort; the model may ignore the constraint, in which case the predicate simply fails.