I’ve been working on Lexa ( https://cerevox.ai/ ) — a document parser that reduces token count in embedding chunks by up to 50% while preserving meaning, making it easier to build high-performing RAG agents.
The idea came from building a personal finance chatbot and hitting token limits fast — especially with PDFs and earnings reports full of dense, noisy tables. Our chunks were bloated with irrelevant data, which killed retrieval quality. So we built Lexa to clean and restructure documents: it clusters context intelligently and removes any spacing, delimiters, or artifacts that don’t add semantic value. The result is cleaner, more compact embeddings — and better answers from your LLM.
Next up, I’m exploring ways to let users fine-tune the parsing logic — using templates, heuristics, or custom rules through UI — to adapt Lexa to their domain-specific needs.