Le Potato
Simple chat interface
Le Potato is a clean, intuitive chat interface for interacting with text-based LLM models. No clutter, just a straightforward way to talk to your AI.
Design
I went for a clean, minimalist UI where components stay modular and reusable. The idea was simple: build something that wouldn't lock the code into this specific project. If myself (or anyone else) wants to repurpose parts for a different tool later, the structure should make that easy, not force a complete rewrite.
Retrieval-Augmented Generation (RAG)
Lets you upload files and pull relevant context into your conversations. Instead of vague responses, your agent actually knows what you're referring to.
Building Better RAG
For solid retrieval performance, I looked at how the big boys handle it and landed on Contextual Retrieval from Anthropic, their benchmarks show clear improvements over traditional RAG approaches.
The rest came down to picking reliable components:
- Embedding model: mixedbread-ai/mxbai-embed-large
- Reranking model: mixedbread-ai/mxbai-rerank-large-v1
Basic Pipeline
With Reranking
Does It Actually Work?
Of course, the results are solid and it's definitely worth building if you want better retrieval without overcomplicating things.
Read Anthropic's full write-up on contextual retrieval
You already know the pain. Smaller models simply couldn't keep up earlier 2024. They'd hallucinate when context got noisy, fail to meaningfully leverage the documents you'd retrieved, or quietly discard the injected context altogether and it left me deep in pipeline debugging, wondering whether anything I made was reaching the model at all.