• 

  ARCHITECTURE

• 

  gallery page includes videos

• 

  dashboard page

Inspiration

• We wanted a simple way for anyone to turn a text prompt into a clear explainer video without manual editing.
• Help learners, educators, and teams go from words to visuals in minutes, especially for STEM concepts that benefit from precise, step-by-step visuals.

What It Does

• You type a text prompt.
  Example: Explain the difference between integration and differentiation.
• The system understands the request, plans an animation, and runs an agentic workflow behind the scenes.
• It renders a polished explainer video and returns a shareable link, so no video editing is required.

How We Built It

Core Idea

• We use the Manim Python library to generate videos directly from code.
  This means animations are produced through precise instructions instead of traditional video editors.
• Existing video-generation models hallucinate or fail at STEM explanations. Code-driven rendering avoids that problem entirely.

Agentic Workflow

1. Preparing the Knowledge Base

• We take the Manim documentation and parse it using Python’s ast module.
• Using hierarchical chunking, we split the docs into meaningful pieces.
• These chunks are embedded and stored in ChromaDB (vector database).

2. Understanding the User Prompt

• A user enters something like: draw a circle and a parabola.
• The enhance prompt agent expands this into k expanded prompts:
  1) Draw a white background. 2) Draw a circle at the center of the screen. 3) Draw a parabola 0.2, 0.2 above center ...

3. Retrieving Relevant Knowledge

• The get chunks agent retrieves top-2 relevant chunks from the vector DB for each prompt step.
  This produces a pool of 2k grounded references.

4. Generating the Animation Code

• The generate code agent get these mapped enhanced prompt with its retrieved chunks.
• It produces valid Python code using Manim.

5. Executing and Delivering the Video

• The execute code agent runs the code in an isolated environment.
• After rendering the MP4 file, it uploads the output to Cloudflare R2 and returns a video URL.
• The frontend receives the link and displays the final video.

Challenges We Ran Into

• Implementing advanced RAG with AST-based hierarchical chunking was difficult.
• Learning the ast library took time.
• Latency remains an issue: videos sometimes take 4–5 minutes to render.
• Accuracy for difficult prompts is still inconsistent.

Accomplishments We're Proud Of

• Successfully deployed the system and made it available on the internet.
• Fixed several frontend issues along the way.
• The approach is fully original this solution does not exist publicly anywhere.

What We Learned

• How to orchestrate multi-step workflows using LangGraph.
• Practical implementation of RAG systems.
• Deep familiarity with Python’s ast module.

What’s Next for vectoraAI

• Reduce end-to-end latency to ~90 seconds.
• Improve accuracy and create custom evals for measurable progress.
• Add a voiceover feature so videos include AI-generated narration.
• Introduce semantic caching with Redis, which will also reduce OpenAI token usage.

Built With

• ast
• cloudflare
• e2b
• fastapi
• langchain
• langgraph
• openai
• supabase