Embed a sentence, send its window.
The Larkspur handbook’s two sections, split by SentenceWindowNodeParser into 17 sentence nodes and embedded with all-MiniLM-L6-v2. Pick a question, move window_size and similarity_top_k, and read what MetadataReplacementPostProcessor hands the model.
- Characters the model reads
- Against the bare sentence
- Answer facts in context
- Sentences sent twice
What the model receives
Retrieved sentence highlighted, answer facts underlined
The handbook, one node per sentence
Cosine similarity, all-MiniLM-L6-v2
| Node | Score | Answer | Sentence |
|---|
Scores come from the real LlamaIndex index and are the same at every window size, because the parser excludes the window from the embedding. The notebook embeds with OpenAI’s text-embedding-3-small, which scores differently. Windows here are a port of the parser’s join, checked against LlamaIndex 0.14.25 on all 30 settings of question, window and k: identical text.