llama-index-jev

# LlamaIndex + TypeSafe Jev

Drop-in LlamaIndex **reranker** and **router** powered by [TypeSafe Jev](https://typesafe.ai): typed `Score` / `Choice` answers, cheap compared to LLM-as-judge — not a Cohere or FlagEmbedding cross-encoder.

[](https://pypi.org/project/llama-index-postprocessor-jev/)
[](https://pypi.org/project/llama-index-selectors-jev/)
[](https://pypi.org/project/llama-index-postprocessor-jev/)
[](https://wiktorb2004.github.io/llama-index-jev/)
[](https://github.com/WiktorB2004/llama-index-jev/actions/workflows/ci.yml)
[](LICENSE)
[](https://pypi.org/project/llama-index-postprocessor-jev/)

Independent community project. **Not** affiliated with TypeSafe or LlamaIndex.

## Install

```bash
pip install llama-index-postprocessor-jev   # JevRerank
pip install llama-index-selectors-jev       # JevSingleSelector, JevMultiSelector
export TYPESAFE_API_KEY=...                 # or OPENROUTER_API_KEY + provider="openrouter"
                                            # or AI_GATEWAY_API_KEY + provider="vercel"
                                            # or CLOUDFLARE_API_TOKEN + CLOUDFLARE_ACCOUNT_ID + provider="cloudflare"
```

## Quickstart

**Rerank** — score each retrieved passage, keep the top `n`:

```python
from llama_index.postprocessor.jev import JevRerank

reranker = JevRerank(top_n=5, mode="score")
# OpenRouter: JevRerank(provider="openrouter", top_n=5, mode="score")
# Vercel:     JevRerank(provider="vercel", top_n=5, mode="score")
# Cloudflare: JevRerank(provider="cloudflare", top_n=5, mode="score")
query_engine = index.as_query_engine(node_postprocessors=[reranker])
```

**Select** — pick which query engine / tool handles the query:

```python
from llama_index.core.query_engine import RouterQueryEngine
from llama_index.selectors.jev import JevSingleS