FactTrainer

FactTrainer is a prototype AI agent training platform. It enables material learning and structured knowledge, while advanced language and reasoning are still developing.

Comment

Inspiration

Many AI assistants today can answer questions well, but they usually rely on large pretrained knowledge or simple document retrieval. We wanted to explore a different approach: can an AI agent learn from user-provided materials and build its own knowledge system?

This idea led to FactTrainer, a platform for training personalized AI agents.

What we built

FactTrainer is an experimental AI agent training platform that allows users to create agents, provide learning materials, and train them through a structured learning pipeline.

The current prototype implements:

  • Material-based learning workflow
  • Teacher agent for organizing knowledge
  • Trainer pipeline for transferring knowledge
  • Structured knowledge representation
  • Knowledge graph storage
  • Basic reasoning and question answering
  • Persistent local agent state

Instead of only storing text answers, FactTrainer converts learning materials into concepts, attributes, relations, rules, and causal links, allowing agents to answer based on learned knowledge.

How it works

The workflow is:

User materials
→ Teacher organizes knowledge
→ Structured knowledge extraction
→ Trainer transfers knowledge
→ Agent builds knowledge graph
→ Agent reasons and answers questions

Current progress

The core learning pipeline is functional.

The prototype can:

  • Create independent agents with isolated knowledge
  • Train agents from uploaded materials
  • Store learned knowledge locally
  • Answer questions based on trained knowledge
  • Preserve agent state after restarting

Challenges and limitations

Building a complete learning agent requires solving several difficult problems, including natural language understanding, knowledge extraction, reasoning, and evaluation.

The current prototype focuses on proving the complete learning workflow. More advanced language understanding, deeper reasoning, and more autonomous knowledge acquisition are still under development.

Future work

Future versions will focus on:

  • Stronger natural language understanding
  • Deeper multi-step reasoning
  • More flexible knowledge extraction
  • Better evaluation of agent learning ability

For the best demonstration experience, please use the provided demo video.

The training process uses Gemini API for Teacher-assisted knowledge extraction. After training, the Agent's learned knowledge and answering ability run locally without requiring an API key.

No API key is included in the repository for security reasons.

Built With

Share this project:

Updates