AI Integration in Fibery: Enhancing Flexibility and Functionality

Two our devs, Vladimir and Max, have recently popped up at the AI Focused Tech Spot in Warsaw. Our guys held talk about AI integration Fibery - this article is an excerpt of the most exciting bits of their presentation.

If you’ve been using Fibery for a while, you probably have a distinct take on how it’s best described. The same applies to us and other users as well:

• Our marketing team describes Fibery as an operating system for companies.
• Our user champions say we create a flexible constructor for business processes.
• Our engineering team thinks we’ve developed yet another flexible no-code tool.

What everyone agrees on is that Fibery is something flexible.

What Does True Flexibility Mean?

Fibery is built on the idea that most business cases for small to medium businesses can be expressed as an open set of objects with special properties.

We call such objects databases.

For instance, the scrum process for software development can be expressed in terms of Releases, Stories, Backlogs, and Sprints:

A sales CRM process can be expressed as a CRM with objects like Accounts, Activities, Opportunities, and Contacts:

So, what makes Fibery’s databases special? They have a set of properties that can be divided into three parts:

• Schema: This defines the data structure, including relations and fields for structured and unstructured data. Relations are powerful because they allow the system to scale and adapt to user needs.
• Formulas and Rules: These are behavioral properties that ensure the data structure follows specific rules.
• Permissions and Views: These are about access and visualization, defining how to represent and manipulate data from a database. Visualization can include tables, boards, lists, timelines, calendars, reports, charts, feeds, maps, documents, whiteboards, and forms. Plus, you can chat with your data using custom chatbots.

Imagine if we weakened this property set by removing relations, rich text fields, and permissions, leaving only basic charts and tables.

You’d get something quite similar to Excel. Fibery takes this idea to the next level by adding new dimensions to the property set and enabling multi-user collaboration. It connects all information within an organization and can replace or integrate data from different tools typically used for tasks like CRM and project management.

Our sales portfolio includes schools, educational organizations, laundries, churches, libraries, and even cannabis farms - further proving the flexibility of Fibery.

Fibery x AI: a Truly Powerful Combo

Schema, Formulas & Views can be represented using JSON (for structured data) and Text (for unstructured data). Both of these things can be generated using modern models.

To improve generation for different tasks, we applied a technique called Prompt Engineering. Prompts vary depending on the task: text generation, creating fields & relations, writing formulas, or generating views. However, the underlying implementation remains the same.

For example, when generating formulas, we provide the model with the name, description, request, and related schema to create the formula.

Creating effective formulas can be challenging, so we use in-context learning for it. Initially, we tried to use predefined examples but found that the model often used fields from the examples instead of the user’s schema. So, we switched to generating examples dynamically based on the incoming user schema.

This approach allows us to avoid hallucination issues and focus solely on user data. It may take a bit longer, but it greatly improves performance since the model has no predefined schemas to rely on.

Embeddings for Search & Clustering

Embeddings are vectors that represent the semantic meaning behind text. They allow us to search for similar meanings and perform mathematical operations on them. In Fibery, we embed user-generated text, such as tasks and feedback, to improve search experience.

When a user query comes in, Fibery uses both keyword search and semantic search. It embeds the user query and searches the vector database, mixing results from both systems to provide the best out of two worlds. Embeddings are also used for clustering, helping to group closely related text chunks together. This is particularly useful for analyzing feedback to identify common themes.

Want More?

To see & hear our guys, check out the entire recording here: