Phoenix Docs Search — Agent-Native Retrieval Demo

How to use

Select a page: click the dropdown and pick one of the 10 indexed pages — results appear automatically
Ranked results: each result shows the source title, section heading, text excerpt, relevance score, and a direct link to the official Phoenix doc page
Scores (0.000–1.000): a keyword relevance score — not a RAG or LLM score. Uses TF-IDF (Term Frequency–Inverse Document Frequency) — a classic information retrieval method that scores how well your query terms match each chunk. TF counts how often a word appears in a chunk. IDF weights rare words higher than common ones. The two are multiplied, then compared using cosine similarity. Common terms like "tracing" score lower (appear in many chunks); specific terms like "playground" or "PXI" score higher (fewer pages). Above 0.05 = relevant. Below = insufficient evidence. The full MCP server uses sentence-transformers for real semantic retrieval — this browser demo uses TF-IDF as a lightweight client-side alternative. See Arize's scikit-learn cosine similarity evaluator and retrieval relevance docs for how Phoenix scores retrieval in production.
Abstention test: try a fake feature (e.g. QuantumBananaTracer) — the system returns no results rather than inventing an answer

🤖 Phoenix Docs Search