lmd ~main

LMD

Library for interacting/managing language models, particularly locally, with native structures for D. Connect to OpenAI, LMStudio, or any OpenAI-compatible endpoint through:

• /v1/chat/completions
• /v1/embeddings
• /v1/models

Built-in support for other providers like Claude and Gemini are planned, but have not yet been implemented. LMD allows for defining new models and endpoints easily and building off of existing endpoints.

Quick Start

import lmd.common.openai;
import lmd.endpoint;

IEndpoint ep = cast(IEndpoint)(new OpenAI("http", "127.0.0.1", 1234));
Response resp = ep.complete!"gpt-oss"("What is 2+2?");
string answer = resp.select!string();

Usage

Endpoints

import lmd.common.openai;

// Local LMStudio
IEndpoint ep = cast(IEndpoint)(new OpenAI("http", "127.0.0.1", 1234));

// OpenAI API
IEndpoint openai = cast(IEndpoint)(new OpenAI("https", "api.openai.com", 443, "your-api-key"));

Completions

import lmd.endpoint;
import lmd.context;

// Simple text
Response resp = ep.complete!"gpt-4"("Explain quantum computing");
string text = resp.select!string();

// With conversation context
Context ctx;
ctx.add(Role.System, "You are a helpful assistant");
ctx.add(Role.User, "What is the capital of France?");
Response resp = ep.complete!"gpt-4"(ctx);

Model Configuration

import lmd.model;

IModel model = ep.fetch("gpt-4");
model.temperature = 0.7;
model.maxTokens = 500;

// Options applied automatically to requests
Response resp = ep.complete!"gpt-4"("Hello!");

Tools

import std.json;
import lmd.tool;

IModel model = ep.fetch("gpt-4");

JSONValue params = parseJSON(`{
    "type": "object",
    "properties": {
        "city": {"type": "string", "desc": "City name"}
    },
    "required": ["city"]
}`);

model.tools().add("get_weather", "Get current weather", params);

Response resp = ep.complete!"gpt-4"("What's the weather in Paris?");
Completion comp = resp.select!Completion();

if (comp.context.messages[0].isTool())
{
    Tool toolCall = comp.context.messages[0].tool();
    string city = toolCall.argument!string("city");
}

Streaming

ResponseStream stream = ep.stream("gpt-4", "Tell me a story");
stream.callback = (Response resp) {
    string text = resp.select!string();
    write(text);
};
// Calling `begin` starts the streaming.
// Callback will be used if it was provided, otherwise you may use `next()`, `last()`, and `collect(size_t)`
stream.begin();

Embeddings

Response resp = ep.embed!"text-embedding-ada-002"("Hello, world!");
float[] vector = resp.select!Embedding().value;

Architecture

The library is endpoint-centric. Endpoints handle API communication, models are lightweight configuration handles, and context is passed as data.

• IEndpoint - Manages HTTP requests and model availability
• IModel - Holds configuration (temperature, tools, etc.)
• Response - Contains Variant data extracted via select!T()
• Context - Conversation messages with roles

Examples

See source/examples/ for working code:

• vibesort.d - Array sorting via LLM
• tool.d - Function calling

Roadmap

• [x] Chat completions
• [x] Embeddings
• [x] Tool/function calling
• [x] Response streaming (lock-free) mostly
• [ ] Tool choice control
• [ ] Image generation
• [ ] Audio endpoints
• [ ] Claude and Gemini providers

License

LMD is licensed under AGPL-3.0.