Build Algebra RAG Assistant
Budget: $250 – $750 USD
I want to put an AI tutor in every algebra student’s pocket. To make that happen, I need you to design and code a Retrieval-Augmented Generation pipeline that can
understand a question, pull the most relevant snippets from my curated algebra textbooks, formula sheets, and problem sets, then explain the answer step-by-step with rock-solid reasoning.
Here is what success looks like from my side:
• A working RAG stack (embeddings → vector database → LLM) wired together with modern tooling such as LangChain, LlamaIndex, or an approach you can justify. FAISS, Pinecone, Weaviate, or similar are fine for the vector layer; OpenAI, Claude, Llama 2/3, or comparable models may drive generation.
• Retrieval tuned for high recall and low latency—students should see an accurate first response in well under a second for typical queries.
• Step-wise explanations that a high-school learner can follow, with all reasoning made explicit and each algebraic transformation clearly spelled out.
• A lightweight interface—REST endpoints are enough, but if you prefer a minimal web front end or ChatGPT-style playground, that’s great as long as it’s clean and documented.
• Configuration files or scripts so I can swap in new algebra content or expand later to calculus and geometry without rewriting code.
• Clear deployment instructions (Docker or similar) plus a short read-through of how to monitor performance and retrain embeddings as the corpus grows.
Acceptance criteria
1. Given a test set of 100 algebra problems I supply, the system returns the correct final answer and an interpretable rationale at ≥90 % accuracy.
2. Average end-to-end response time ≤800 ms on a single GPU or cloud instance I specify.
3. All code, environment files, and a concise README delivered in a private Git repo.
If you’ve built RAG systems before—or have a clever idea to push accuracy beyond standard chunk-and-vector methods—I’d love to see a brief note on your approach along with any relevant repos or demos.
Here is what success looks like from my side:
• A working RAG stack (embeddings → vector database → LLM) wired together with modern tooling such as LangChain, LlamaIndex, or an approach you can justify. FAISS, Pinecone, Weaviate, or similar are fine for the vector layer; OpenAI, Claude, Llama 2/3, or comparable models may drive generation.
• Retrieval tuned for high recall and low latency—students should see an accurate first response in well under a second for typical queries.
• Step-wise explanations that a high-school learner can follow, with all reasoning made explicit and each algebraic transformation clearly spelled out.
• A lightweight interface—REST endpoints are enough, but if you prefer a minimal web front end or ChatGPT-style playground, that’s great as long as it’s clean and documented.
• Configuration files or scripts so I can swap in new algebra content or expand later to calculus and geometry without rewriting code.
• Clear deployment instructions (Docker or similar) plus a short read-through of how to monitor performance and retrain embeddings as the corpus grows.
Acceptance criteria
1. Given a test set of 100 algebra problems I supply, the system returns the correct final answer and an interpretable rationale at ≥90 % accuracy.
2. Average end-to-end response time ≤800 ms on a single GPU or cloud instance I specify.
3. All code, environment files, and a concise README delivered in a private Git repo.
If you’ve built RAG systems before—or have a clever idea to push accuracy beyond standard chunk-and-vector methods—I’d love to see a brief note on your approach along with any relevant repos or demos.