PROMPTAGATOR

A few-shot approach to domain-specific retrieval from Google Research: prompt a very large language model with a handful of labeled examples, generate synthetic in-domain training data, and train a retriever and reranker on it. Published as Promptagator: Few-shot Dense Retrieval From 8 Examples (arXiv:2209.11755).

It is the reference point the Vespa zero-shot series measures itself against, in both directions — first as the stronger few-shot method that a zero-shot pipeline cannot quite reach, then as the method a much smaller generator matches.


Position Relative to Zero-Shot Hybrid

From Improving Zero-Shot Ranking with Vespa Hybrid Search - part two, averaged over the subset of BEIR datasets PROMPTAGATOR reports:

ModelAverage nDCG@10
Vespa hybrid (zero-shot, no in-domain data)0.456
PROMPTAGATOR (dense retriever)0.478
PROMPTAGATOR (cross-encoder)0.528

Per-dataset figures available from that article:

DatasetVespa HybridPROMPTAGATOR (dense)PROMPTAGATOR (cross-encoder)
TREC-COVID0.7500.7560.762
FiQA-20180.2920.4620.494
ArguAna0.4040.5940.630
HotpotQA0.6320.6140.736

The reported averages do not reconcile with these four rows alone, so the comparison spans more datasets than are captured here.

The comparison is not apples to apples, and the article says so: PROMPTAGATOR uses in-domain synthetic data and is therefore few-shot, not zero-shot. Its cross-encoder also re-ranks the top 200 and depends on billion-parameter LLM inference for the generation step. The zero-shot hybrid’s claim is efficiency — CPU-only serving, sub-60 ms latency, no per-domain training step — rather than higher quality.

Then Beaten by a 3B Generator

A month later, Improving Search Ranking with Few-Shot Prompting of LLMs applied the same overall recipe to TREC-COVID with a 3B generator (FLAN-T5 xl) instead of PROMPTAGATOR’s 137B FLAN model:

ModelTREC-COVID nDCG@10
Cross-encoder on 3B-generated synthetic data80.2
PROMPTAGATOR (137B FLAN)76.2

A roughly 45× smaller generator produced better downstream ranking on this dataset. The inference to draw is narrow — one dataset, one comparison — but it points at prompt design and Consistency Filtering mattering more than raw generator scale.

Datasets

  • BEIR — the evaluation suite for the comparison
  • TREC-COVID — the single dataset where a 3B generator beat it

Articles

Source