Generative Modules

Generative Modules enable you to integrate retrieval-augmented generation (RAG) and LLM-powered capabilities directly into your Unbody projects. These modules allow your applications to produce intelligent, context-aware outputs by combining structured data with advanced AI models.

What are Generative Modules?

Generative Modules are configurations that enable features like:

• Question Answering: Combine vectorized data with generative responses.
• Content Creation: Generate new content based on stored knowledge.
• Dynamic Responses: Tailor application outputs to specific user queries.

They are flexible and model-agnostic, supporting multiple providers such as OpenAI, Cohere, and others.

How They Work

Generative Modules work as part of your project’s pipeline:

1. Data Retrieval: Fetches relevant knowledge from the vector database.
2. Context Injection: Combines retrieved data with a query.
3. Generation: Uses an LLM to produce structured or unstructured outputs.

Configuring Generative Modules

You can configure Generative Modules via the Dashboard or the Admin API.

1. Dashboard Setup

To enable Generative Modules:

1. Go to your Project Settings in the Dashboard.
2. Navigate to the Generative Modules section.
3. Select and configure a model provider (e.g., OpenAI GPT-4 or Cohere Command).
4. Save your settings.

2. Admin API Configuration

You can programmatically configure Generative Modules using the Admin API.

Example: Setting Up a Generative Module

import { UnbodyAdmin, Generative } from 'unbody/admin'
 
// Initialize the Admin API
const admin = new UnbodyAdmin({
  auth: {
    username: '[admin-key-id]',
    password: '[admin-key-secret]',
  },
})
 
// Fetch the project
const project = await admin.projects.get({ id: '[project-id]' })
 
// Configure a generative module
const generative = new Generative(Generative.OpenAI.GPT4)
console.log(generative.options.model) // "gpt-4"
 
// Apply the generative module to the project settings
const settings = project.settings.set(generative)
await project.updateSettings(settings)
 
console.log(`Added Generative Module to project: ${project.name}`)
 

Available Generative Models

Unbody supports multiple providers and models. Below is a list of supported options:

For a full list of available models, refer to the Model Reference Documentation.

Advanced Settings

Context Injection Strategies

Define how retrieved data is injected into prompts for generation:

• Full Context: Inject all retrieved results.
• Condensed Context: Use summarization or filtering to reduce context size.

Prompt Tuning

Customize prompts to fine-tune model outputs. Use placeholders like:

• {context}: Injects retrieved data.
• {query}: Represents the user’s input query.

Example:

"Based on the following context: {context}, answer the question: {query}"

Temperature

Adjusts creativity levels. Lower values produce deterministic results, while higher values encourage creativity.

Best Practices

1. Model Selection:
   • Use GPT-4 or GPT-3.5 Turbo for general-purpose tasks.
   • Opt for Cohere or Mistral models for domain-specific needs.
2. Context Optimization:
   • Limit the context size for faster generation.
   • Use summarization in preprocessing pipelines to refine inputs.
3. Error Handling:
   • Implement fallback models for improved reliability.