Models

INTELLITHING hosts its own models locally, although they are Hugging Face-compatible models originating from the Hugging Face ecosystem. For compatibility and configuration purposes, the INTELLITHING platform currently requires an IntelliConfig file. As a result, the team continuously curates and hosts the most reliable and well-known models in the model directory.

INTELLITHING also supports the Ollama framework, which enables powerful local workflows and seamless model deployment.

---

🔑 Key Concepts

There are two categories of models supported in INTELLITHING:

1. Large Language Models (LLMs)
2. TML (Traditional Machine Learning)

INTELLITHING’s hybrid philosophy is grounded in enterprise-grade reliability. Solely relying on LLMs for decision-making is not recommended in critical or regulated environments. Therefore, TML models are treated as tools within LLM workflows—just like other blocks or utilities. This allows for precision decision-making, auditability, and enhanced explainability when required.

INTELLITHING automates the integration, input, and output management between LLMs and TML models within your workflows.

⚠️ Unlike pre-trained LLMs, TML models require training on your own data. You can use real datasets or leverage INTELLITHING’s powerful data synthesizer to generate synthetic data. For details, refer to the "Train a Model" section.

📘 Key Definitions

1. LLM (Large Language Model):
   Pre-trained transformer-based models used within the block editor. In INTELLITHING workflows, LLMs primarily serve as orchestrators—integrating multiple tools and blocks—rather than being used for critical or final decision-making.

2. TML (Traditional Machine Learning):
   Explainable models trained on structured, domain-specific data. TML models are optimized for narrowly scoped tasks, require feature engineering, and offer fine control over input-output behavior. They are evaluated using standard metrics like error rates, correlation scores, and confusion matrices—offering transparency and interpretability.

INTELLITHING’s AutoML engine automates the entire ML pipeline—from data preprocessing and model selection to hyperparameter optimization, model explainability, and resource management. Once trained, the TML model is available as a drag-and-drop block in the block editor.

⚙️ LLM Configuration

Just like any other block, an LLM block can be configured by clicking on it within the block editor. This opens a configuration prompt with several customizable options.

ℹ️ If you’re using INTELLITHING v4.0 or later, instance selection (CPU/GPU) is done via the project card under the deployment & utilities section.

Configuration Options

1. GPU/CPU Selection
   Define the number and type of compute resources to use for this deployment.

2. Chat Engine Toggle
   Select this if you are building a chat application instead of a general-purpose workflow automation. When enabled, INTELLITHING will automatically attach the Chat Engine and Chat Memory blocks to your deployment.

a. Chat Engine

A high-level interface enabling multi-turn conversations with your data. It supports context-aware interactions, allowing users to follow up on previous queries. It works by managing message flow and integrating with LLMs and data sources to generate coherent, contextual responses.

b. Chat Memory

Manages the history of interactions within a session. It ensures that previous messages are stored and retrievable, allowing the chat engine to incorporate past context into its replies.
Memory types like ChatMemoryBuffer allow token-limited history retention and message summarization to manage long conversations.

⚙️ TML Configuration

Once a Traditional Machine Learning model has been trained, you can drag and drop it as a block in the editor. Clicking on the block opens a configuration prompt that dynamically displays the model’s input and output variables along with editable description fields.

INTELLITHING’s auto-formatting system uses this metadata to intelligently route inputs and outputs—even in complex, chat-heavy environments where data may be unstructured or semi-structured.