Built this after debugging too many RAG systems where the chunks were the problem, not the model or retriever.
ChunkScore scores chunks across 5 weighted signals before they hit your vector database:
- Length adequacy
- Completeness (no orphaned context)
- Signal clarity (noise ratio)
- Semantic density
- Context sufficiency
Works on any chunks from any source — LangChain, LlamaIndex, Chonkie, RAGFlow, or raw JSON arrays from inherited vector DBs.
What makes it different from RAGAS/DeepEval:
Those evaluate after generation. ChunkScore catches quality failures before embedding — so you’re not paying to store and retrieve bad data in the first place.
Free features:
- Upload TXT, MD, CSV, PDF, DOCX
- 3 chunking strategies with side-by-side comparison
- Import pre-chunked JSON arrays for quality audit
- Export to Pinecone, Weaviate, Chroma, JSON, CSV
- Runs entirely in your browser — data never leaves your machine
Happy to answer questions about the scoring methodology or chunking strategy comparisons.