ollamer

What is ~~love~~ ollamer?

Minimalistic wrapper for ollama for your experiments with building workflows locally.

Build your prompts with JSON output and use them as regular async functions.

Usage example

Create folders with your prompts:

src
  tasks
    check_relevance
      index.ts
      article.hbs
  index.ts

Create ollama task.

// src/tasks/summarize/index.ts
import joi from 'joi';
import { resolve } from 'node:path';
import { createTask } from 'ollamer';

// argument for our function
export type Input = {
  title: string;
  article: string;
};

// expected output format for out function
export type Relevance = {
  related: boolean;
  confidence: number;
  reason: string;
};

export default createTask<Input, Relevance>({
  model: 'llama3.2',

  name: 'check_relevance',

  options: {
    // Ollama's options
    temperature: 0.1,
  },

  retries: 1,

  timeout: 30 * 1000,

  // Use `format` field to create JSON typing (it will be verified)
  format: joi.object<RelevanceResult>({
    related: joi.boolean()
      .required(),
    confidence: joi.number()
      .min(0)
      .max(1)
      .required(),
    reason: joi.string()
      .required(),
  }),

  templates: [
    {
      role: 'system',
      filename: resolve(import.meta.dir, './system.hbs'),
    },
    {
      role: 'user',
      filename: resolve(import.meta.dir, './article.hbs'),
    },
  ],
});

Use this task as it is usual asynchronous function:

// src/index.ts
import summarizeTask, type { Input, Relevance } from './tasks/summarize';
import type { Result } from 'ollamer';

const title = 'Using Ollama never was so convenient!';
const article = ` ... your articles are beautiful ... `;

const result: Result<Input, Relevance> = await summarizeTask.run({
  title,
  article,
});

console.log(result.output); // type defined with your Joi schema

That's it. Read package's JSDocs for more details.