---
license: mit
task_categories:
- question-answering
- sentence-similarity
language:
- en
tags:
- retrieval
- rag
- reranking
- information-retrieval
pretty_name: Tiny RAG with Reranking (eval set + chunk sample)
size_categories:
- n<1K
---

# tiny-rag-with-reranking

A tiny retrieval / question-answering evaluation set plus a sample of a chunked
public-domain corpus, used by the
[tiny-rag-with-reranking](https://github.com/narinzar/tiny-rag-with-reranking)
RAG pipeline (bi-encoder retrieval + cross-encoder reranking, with a chunk-size
sweep).

## Task

Question answering / passage retrieval. Each query carries one or more short
answer substrings; a retrieved passage is judged relevant if it contains any
answer substring (case- and whitespace-insensitive). This substring labeling
keeps relevance valid regardless of how the corpus is chunked.

## Files

- `qa.json` — the evaluation set: 12 queries, each with a `question`, one or
  more `answers` (relevant-passage substrings), and the `source` document.
- `chunks_sample.json` — a 60-chunk sample (of 902 total in
  the full build) of the chunked corpus, each with `doc`, character offsets
  `start`/`end`, and `text`. Produced by the adaptive chunker at target size 128.

## Generation method

The corpus is public-domain plaintext from Project Gutenberg (Alice's Adventures
in Wonderland, The Time Machine, A Study in Scarlet), with license headers and
footers stripped, then chunked with an adaptive sentence-merging strategy. The
query/relevant-passage labels in `qa.json` are hand-written: canonical phrases
drawn verbatim from the corpus so they stay stable across chunkings.

## Measured results (small-scale benchmark, single RTX 5090)

Bi-encoder `all-MiniLM-L6-v2` retrieval vs adding cross-encoder
`ms-marco-MiniLM-L-6-v2` reranking, retrieve top-20, metric @k=5:

| stage                  | precision@5 | recall@5 |
|------------------------|-------------|----------|
| bi-encoder only        | 0.1333      | 0.0650   |
| + cross-encoder rerank | 0.2000      | 0.1394   |

Chunk-size sweep best size by precision@5: 512.

## License

MIT. Underlying source texts are public domain (Project Gutenberg).