OpenAI Provider

@flowrag/provider-openai provides embeddings, entity extraction, and reranking via the OpenAI API — or any OpenAI-compatible endpoint.

Installation

npm install @flowrag/provider-openai

Configuration

Set your API key:

export OPENAI_API_KEY=your-key

Embedder

import { OpenAIEmbedder } from '@flowrag/provider-openai';

const embedder = new OpenAIEmbedder({
  model: 'text-embedding-3-small', // default
  dimensions: 1536,                // default
});

embedder.dimensions; // 1536

Available models:

Model	Default Dimensions
text-embedding-3-small	1536
text-embedding-3-large	3072

The dimensions parameter can be customized — OpenAI supports shortening embeddings.

Extractor

import { OpenAIExtractor } from '@flowrag/provider-openai';

const extractor = new OpenAIExtractor({
  model: 'gpt-5-mini', // default
});

Uses JSON mode (response_format: { type: 'json_object' }) for reliable structured output.

Reranker

import { OpenAIReranker } from '@flowrag/provider-openai';

const reranker = new OpenAIReranker();

LLM-based relevance scoring — each result is evaluated against the query.

OpenAI-Compatible Endpoints

All classes accept a baseURL option, so you can point them at any OpenAI-compatible API:

// Ollama (local, free)
const embedder = new OpenAIEmbedder({
  baseURL: 'http://localhost:11434/v1',
  model: 'nomic-embed-text',
  dimensions: 768,
});

// Azure OpenAI
const extractor = new OpenAIExtractor({
  baseURL: 'https://my-company.openai.azure.com',
  apiKey: process.env.AZURE_OPENAI_KEY,
  model: 'gpt-4o',
});

// Together, Groq, Mistral, vLLM, LiteLLM...
const reranker = new OpenAIReranker({
  baseURL: 'https://api.together.xyz/v1',
  apiKey: process.env.TOGETHER_API_KEY,
  model: 'meta-llama/Llama-3-70b',
});

This works with any service that implements the OpenAI API format, including:

• Ollama — local models, no API key needed
• Azure OpenAI
• vLLM — self-hosted inference
• LiteLLM — unified proxy for 100+ providers
• Together AI
• Groq
• Mistral

Model Constants

Well-known model IDs are exported as constants:

import { OpenAIEmbeddingModels, OpenAILLMModels } from '@flowrag/provider-openai';

OpenAIEmbeddingModels.TEXT_EMBEDDING_3_SMALL; // 'text-embedding-3-small'
OpenAIEmbeddingModels.TEXT_EMBEDDING_3_LARGE; // 'text-embedding-3-large'

OpenAILLMModels.GPT_5_MINI; // 'gpt-5-mini'
OpenAILLMModels.GPT_5;      // 'gpt-5'
OpenAILLMModels.GPT_5_2;    // 'gpt-5.2'

All model parameters accept any string, so you can use newer models as they become available.

Full Example

import { OpenAIEmbedder, OpenAIExtractor, OpenAIReranker } from '@flowrag/provider-openai';

const rag = createFlowRAG({
  schema,
  storage: { kv, vector, graph },
  embedder: new OpenAIEmbedder(),
  extractor: new OpenAIExtractor(),
  reranker: new OpenAIReranker(), // optional
});