RAG & Memory

MovingLines uses Retrieval-Augmented Generation (RAG) to ensure that the AI has access to up-to-date, stable, and high-quality Manim code examples.

The Problem with LLMs and Manim

Standard LLMs often lean on outdated Manim syntax or hallucinate methods because they were trained on a mix of legacy and community editions.

Our Solution: Pinecone Vector Search

We maintain a curated database of "Golden Examples" — Manim scripts known to be bug-free and visually stunning.

How it Works:

1. Semantic Search: When you enter a prompt like "show me a rolling circle", we perform a vector search in Pinecone.
2. Context Injection: The top 5 most relevant code snippets are retrieved.
3. Prompt Augmentation: These snippets are injected into the Generator's system prompt as "Reference Examples."

Benefits of RAG

• Syntax Stability: The AI mimics the modern syntax found in our verified examples.
• Visual Variety: By including diverse examples (3D, Graphing, LaTeX), the AI learns more complex animation patterns.
• Improved Reliability: Significantly reduces the number of "first-attempt" rendering failures.

Seeding the Database

Developers can add new examples to the RAG database using our built-in utility:

cd backend
python -m scripts.seed_pinecone

This script reads from a local collection of high-quality .py files and upserts their embeddings to Pinecone.