Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
RAG-Fusion: multi-query generation + Reciprocal Rank Fusion for better retrieval-augmented generation. Includes evaluation harness with NFCorpus/BEIR.
| Date | Stars |
|---|---|
| 2026-07-24 | 946 |
| 2026-07-25 | 947 |
| 2026-07-28 | 947 |
| 2026-07-30 | 947 |
| 2026-07-31 | 948 |
| 2026-08-06 | 948 |
Today
— stars today
This week
+1 stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.11%/day
# RAG-Fusion: The Next Frontier of Search Technology
## Overview
RAG-Fusion is a search methodology that aims to bridge the gap between traditional search paradigms and the multifaceted dimensions of human queries. Where Retrieval Augmented Generation (RAG) fuses vector search with generative models, RAG-Fusion goes a step further — employing multiple query generation and Reciprocal Rank Fusion to re-rank search results. The aim is to surface relevant material a single phrasing of the query would miss, particularly when the user's vocabulary doesn't match how the corpus is indexed.
For the full story behind the approach, see the article: [Forget RAG, the Future is RAG-Fusion](https://adrianraudaschl.com/blog/forget-rag-the-future-is-rag-fusion/).
> **Where this technique fits, in one line:** Properly-configured RAG-Fusion (`hybrid_diverse+rerank`: BM25 + vector × LLM rewrites, fused via RRF, then cross-encoder reranked) produces measurably better retrieval rankings *and* better generated answers than baseline retrieval — at proper sample sizes with confidence intervals, on every difficulty bucket, even with a strong reranker. **The vector-only fusion variant is a different story** — it's roughly a wash on average and net-negative on rich queries at the answer level. If you deploy fusion, deploy the hybrid variant.
>
> Detailed empirical writeup — n=200 paired-bootstrap CIs, three rerankers, six fusion variants, end-to-end LLM-judge answer eval, including a replication of arXiv [2603.02153v1](https://arxiv.org/html/2603.02153v1) — lives in [`experiments/arxiv-2603-02153-replication/`](./experiments/arxiv-2603-02153-replication/README.md).
## How It Works
```mermaid
flowchart TD
Q[Original Query] --> L[LLM generates multiple queries]
L --> V1[Vector Search 1]
L --> V2[Vector Search 2]
L --> VN[Vector Search N]
V1 --> RRF[Reciprocal Rank Fusion]
V2 --> RRF
VN --> RRF
RRF --> OUT[Re-ranked Docs]
```
1. **Query Generation** — Takes a user's query and uses OpenAI's GPT to generate multiple search query variations that capture different facets of the original intent.
2. **Vector Search** — Conducts vector-based searches using ChromaDB on each query, casting a wider net across the document space.
3. **Reciprocal Rank Fusion** — Combines the ranked results from all searches, boosting documents that appear consistently across multiple query perspectives.
4. **Output Generation** — Produces a final re-ranked list of documents, optionally synthesised into a natural language answer via LLM.
## When to use RAG-Fusion
The technique earns its compute when three conditions hold:
1. **Terminology mismatch between user queries and indexed text** (lay vs technical names, jargon, paraphrase).
2. **Recall matters more than precision** — missing a relevant document is more costly than including a marginal one.
3. **The downstream consumer can handle topically-broad context** — either a strong synthesis LLM, or a UI that surfaces multiple candidates rather than one canonical answer.
Strong-fit examples:
- Academic / scientific literature search, biomedical research
- Patent prior-art search, legal e-discovery, regulatory review
- Long-tail e-commerce ("phone holder thing for car" → "magnetic vent mount")
- Cold-start retrieval over specialist corpora the embedding model hasn't seen
- Exploratory / "show me what's out there" workflows
Poor-fit examples:
- FAQ chatbots and curated customer-support knowledge bases
- Latency-critical retrieval (voice, autocomplete, sub-second-p95 chat)
- High-volume / margin-thin consumer search
- Code or identifier search (precision-dominated)
- Structured data, knowledge graphs, SQL-backed retrieval
For mixed workloads — most production retrieval — the right pattern is **adaptive routing**: run baseline+rerank on every query, fire fusion only when a cheap weakness signal trips. This captures the long-tail wins, eliminates the regression cases on easy queries, and pays forExcerpt of 11,898 characters
Read on GitHub
14
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:2867205674f6f607, topic:rag, topic:retrieval-augmented-generation, desc:retrieval-augmented generation