Examples

September 18, 2026 · View on GitHub

Live walkthroughs of the Jev reranker and selector against OpenRouter's Decisions API. Retrieval and routing quality is the point; embeddings and the unused router summarizer are local mocks so you do not also need an OpenAI key.

export OPENROUTER_API_KEY=...
uv run python examples/basic_rerank.py
uv run python examples/basic_selector.py
uv run python examples/router_query_engine.py

Outputs below were captured with provider="openrouter" and timeout_s=30. Scores and choices can move slightly between runs.

basic_rerank.py

Five in-memory passages, mock embeddings (every cosine is 1.000), then JevRerank(top_n=3, mode="score") on Where is the Eiffel Tower?.

Score mode is a 0–3 rubric (0 off-topic … 3 fully answers), not cosine similarity. Jev promotes the Eiffel Tower sentence to the top and drops the photosynthesis / Python noise.

Before JevRerank:
  1.000  Paris is the capital of France.
  1.000  Photosynthesis converts light energy into chemical energy.
  1.000  The Louvre is the world's largest art museum, in Paris.
  1.000  The Eiffel Tower is a wrought-iron tower in Paris, France.
  1.000  Python is a high-level programming language.

After JevRerank:
  3.000  The Eiffel Tower is a wrought-iron tower in Paris, France.
  0.910  Paris is the capital of France.
  0.690  The Louvre is the world's largest art museum, in Paris.

basic_selector.py

JevSingleSelector over two toy tools (weather vs docs) with no RouterQueryEngine. Query: Will it rain in Paris tomorrow?

query:      Will it rain in Paris tomorrow?
index:      0
reason:     Jev selected 'weather' (confidence=1.00)
chose:      weather

router_query_engine.py

Same two tools, wired through RouterQueryEngine. Jev picks a route, then the chosen engine returns a fixed string.

RouterQueryEngine always constructs an LLM for its multi-route summarizer even when the selector is single-choice. This file passes MockLLM() so that constructor does not try to import llama-index-llms-openai. verbose=True prints the selection before the answer.

Selecting query engine 0: Jev selected 'weather' (confidence=1.00).
Forecast: light rain in Paris tomorrow.