How It Works
📄
1. Load Docs
Paste any text documents
✂️
2. Chunk
Split into overlapping segments
⚡
3. BM25 Search
Okapi BM25 lexical retrieval
🤖
4. Generate
Claude Haiku with source citations
Documents
No documents yet.
Add documents and run a query to see the RAG pipeline in action.
Under the hood:
Retrieval: Okapi BM25 (k₁=1.5, b=0.75) — the same algorithm powering Elasticsearch and Apache Lucene. Production systems often layer neural reranking (cross-encoders) on top of BM25.
Chunking: Paragraph-aware splitting with character-level fallback, ~400-char chunks. Overlap prevents context loss at boundaries.
Generation: Claude Haiku 3.5 with source-citation-enforced system prompt. Zero hallucination risk from knowledge not present in retrieved chunks.