Vendi-RAG

Adaptively trading off diversity and quality significantly improves retrieval-augmented generation with LLMs

Accepted to EMNLP 2026 Findings.

Similarity search scores every document on its own, so what comes back is usually several restatements of the same fact. Multi-hop questions need the opposite: a few documents that say different things. Vendi-RAG scores the retrieved set, trading query relevance against set-level diversity, and puts that behind one knob.

At s = 0 the retriever takes a slice off the right-hand edge: eight documents, fewer than four distinct facts between them, and the evidence missed. Raising s fans the selection out until the evidence is in context. Reproduce it with vendirag demo --gif out.gif.

---

Install

pip install vendirag

The retriever, the Vendi Score, the pipeline, the toy corpus, and the offline backend all run on numpy alone. Everything else is opt-in:

pip install "vendirag[openai]"      # OpenAI models
pip install "vendirag[anthropic]"   # Claude models
pip install "vendirag[embeddings]"  # sentence-transformers encoders
pip install "vendirag[viz]"         # plots and the animation above
pip install "vendirag[research]"    # everything needed to rerun the paper

The retriever, on its own

Index some text, get a set back. No LLM, no vector store, no API key.

from vendirag import VendiRetriever

retriever = VendiRetriever.from_texts(documents, s=0.8, k=5)
docs = retriever.retrieve("Which instrument did the expedition carry?")

s is the only knob: 0 is pure relevance (identical to top-k similarity search), 1 ignores relevance within the pool entirely, and 0.8 is the paper's default. Ask for the diagnostics when you want to see what happened:

result = retriever.retrieve_details("Which instrument did the expedition carry?")
result.vendi_score      # effective number of unique documents in the set
result.mean_similarity  # mean query-document cosine similarity
result.documents        # the selected documents, in selection order

VendiRetriever uses sentence-transformers when it is installed and a built-in deterministic hashing embedder otherwise. Anything with an encode(texts) -> (n, d) method works:

from vendirag import SentenceTransformerEmbedder

retriever = VendiRetriever(
    embedder=SentenceTransformerEmbedder("all-mpnet-base-v2"),
    s=0.8, k=5, candidate_pool=50,
)

Already have a vector store?

Keep it. VendiReranker adds only the selection step, so the corpus stays wherever it lives — Chroma, FAISS, Pinecone, Elasticsearch, pgvector.

from vendirag import VendiReranker

def candidates(query, n):
    """Return (documents, doc_embeddings, query_embedding)."""
    hits = my_index.search(query, n)
    return hits.docs, hits.vectors, my_encoder(query)

reranker = VendiReranker(candidates, s=0.8, k=5, candidate_pool=50)
docs = reranker.retrieve("who led the expedition?")

Or reach for the selection function directly, if you would rather own the loop:

from vendirag import vendi_rerank

docs = vendi_rerank(query_embedding, pool_embeddings, pool_documents, k=5, s=0.8)

The full pipeline

Retrieval is half the method. The other half is a loop that treats s as something to steer rather than tune — when the answer is good, s falls and retrieval consolidates; when it is poor, s rises and retrieval fans out to find what is missing.

from vendirag import VendiRAG, OpenAILLM

rag = VendiRAG(retriever, llm=OpenAILLM("gpt-4o-mini"))
result = rag.answer("In which town is the observatory that houses the instrument?")

print(result.answer)              # best answer across all iterations
print(result.quality)             # judge score in [0, 1]
print(result.s_trajectory)        # how the diversity weight moved
print(result.documents)           # the evidence behind the answer

Claude works the same way, and the judge can be a different, cheaper model — the quality target is relative to the running best rather than absolutely calibrated, so the judge only has to rank, not score:

from vendirag import AnthropicLLM

rag = VendiRAG(
    retriever,
    llm=AnthropicLLM("claude-sonnet-5"),
    judge_llm=OpenAILLM("gpt-4o-mini"),
)

Anything callable is accepted, so a local model or a LangChain runnable drops straight in:

rag = VendiRAG(retriever, llm=lambda prompt: my_local_model(prompt))

Algorithm 1

Initialize  q₁ ← q,  s₁ ← 0.8,  best ← 0
for i = 1 … N:
    C   ← TopK-Similar(qᵢ, 𝒦, |C|)              candidate pool
    Dᵢ  ← GreedyVRS(C, sᵢ, k)                   Eq. 2, below
    rᵢ  ← GenerateCoT(qᵢ, Dᵢ)
    âᵢ  ← GenerateAnswer(q, Dᵢ, r₁…ᵢ)           the ORIGINAL question
    Qᵢ  ← LLM-Judge(âᵢ, q, Dᵢ)                  mean(coherence, relevance, alignment) / 10
    if Qᵢ > best:  â*, best ← âᵢ, Qᵢ
    if Qᵢ ≥ τ:     return â*
    s_target ← clip(1 − Qᵢ / max(best, ε), 0, 1)
    sᵢ₊₁     ← β·sᵢ + (1−β)·s_target            Eq. 3, an EMA
    qᵢ₊₁     ← RefineQuery(qᵢ, âᵢ, rᵢ)
return â*

The selection objective is the Vendi Retrieval Score,

VRS(D) = s · ṼS(D) + (1 − s) · S̃S(q, D)

where ṼS is the Vendi Score of the set — the exponential of the Shannon entropy of the eigenvalues of its similarity kernel, interpretable as the effective number of unique documents — rescaled to [0, 1], and S̃S is the mean query-document cosine similarity min-max normalized over the candidate pool. Documents are added greedily, since maximizing over all k-subsets is intractable.

Cost is O(k · |C| · k³): the eigendecomposition is on the selected subset, of size at most k, never on the corpus.

Configuration

Argument	Default	Meaning
initial_s	0.8	s₁, the starting diversity weight
k_docs	5	k, documents selected per iteration
k_candidates	50	|C|, candidate pool size
max_iterations	5	N, iteration cap
quality_threshold	0.85	τ, the early-stop threshold
beta	0.3	β, EMA momentum
dynamic_s	True	False pins s — the fixed-s variant
early_stopping	True	False always runs N iterations
refine_query	True	False re-retrieves with the original question
judge_llm	None	separate judge model; defaults to the generator
on_iteration	None	callback fired per iteration, for logging or progress

The paper recommends fixed s = 0.8 as the default configuration: s is the substantive knob and a well-chosen constant captures most of the gain at no extra inference cost. Adapting it online is a second-order refinement whose benefit concentrates on hard queries.

rag = VendiRAG(retriever, llm=..., initial_s=0.8, dynamic_s=False)

Just the score

The Vendi Score stands alone as a way to measure how redundant any retriever's output is:

from vendirag import vendi_score

vendi_score(embeddings)   # in [1, n]: 1 if identical, n if mutually orthogonal

Try it

A synthetic multi-hop benchmark ships with the package and runs in seconds with no API key, no model download, and no network:

vendirag demo                              # the numbers below
vendirag demo --gif assets/demo.gif        # and the animation
python examples/toy_demo.py                # the same, annotated

Or work through examples/vendi_rag_toy_demo.ipynb, which builds the whole argument step by step with plots.

The corpus is a fictional world of Arctic expeditions. Answering a question means chaining four facts that live in four separate documents, and around each chain sits a thicket of near-duplicate documents repeating the expedition's name — the way a real corpus accumulates coverage of a notable entity.

Retrieval (k = 8, |C| = 50, 60 questions, deterministic hashing embedder):

method	evidence found	effectively unique docs
top-k similarity	7%	3.7 / 8
MMR (λ = 0.5)	73%	7.4 / 8
MMR (λ = 0.7)	78%	7.3 / 8
Vendi retrieval (s = 0.8)	80%	7.4 / 8
Vendi retrieval (s = 1.0)	57%	7.5 / 8

End to end, using the built-in model-free reader so the run is free and deterministic:

configuration	exact match
single-shot, top-k	0%
single-shot, Vendi s = 0.8	2%
iterative, top-k	13%
Vendi-RAG (fixed s = 0.8)	78%
Vendi-RAG (dynamic s)	73%

Neither half is sufficient alone: one diverse retrieval cannot reach hop 3, and iterating over a redundant set just re-reads the same thicket.

These are synthetic numbers illustrating the mechanism, not benchmark results — the corpus was built to exhibit the failure mode. They are reproducible, since both the corpus and the embedder are deterministic. For real-data results see the paper, or research/ to rerun them.

Reproducing the paper

HotpotQA, MuSiQue, and 2WikiMultiHopQA, with the Chroma index and scoring harness: see research/README.md.

pip install -e ".[research]"
bash research/download/processed_data.sh
python research/ingestion.py --dataset hotpotqa --build
python research/run_benchmark.py --dataset hotpotqa --model gpt-4o-mini
python research/evaluation/cal_f1_em.py --predictions results/<file>.csv

Layout

vendirag/            the library
  vendi.py           Vendi Score primitives          (numpy only)
  retriever.py       VRS selection, VendiRetriever, VendiReranker
  pipeline.py        VendiRAG — Algorithm 1
  llm.py             OpenAI / Anthropic adapters, prompted and offline backends
  embeddings.py      hashing and sentence-transformers embedders
  prompts.py         the paper's prompts, including the exact judge prompt
  toy.py             the synthetic multi-hop corpus and its metrics
  viz.py             plots and the animation above
  cli.py             the `vendirag` command
examples/            runnable demo script and notebook
research/            paper reproduction: ingestion, benchmarks, baselines, scoring
tests/               pytest suite

Citation

@misc{rezaei2025vendiragadaptivelytradingoffdiversity,
      title={Vendi-RAG: Adaptively Trading-Off Diversity And Quality Significantly
             Improves Retrieval Augmented Generation With LLMs},
      author={Mohammad Reza Rezaei and Adji Bousso Dieng},
      year={2025},
      eprint={2502.11228},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.11228},
}

The Vendi Score was introduced in:

@article{friedman2023vendi,
  title={The Vendi Score: A Diversity Evaluation Metric for Machine Learning},
  author={Friedman, Dan and Dieng, Adji Bousso},
  journal={Transactions on Machine Learning Research},
  year={2023},
}

License

MIT © Mohammad Reza Rezaei, Adji Bousso Dieng