Retrieval Augmentation
123 beliefs (123 IN, 0 OUT)
The Retrieval Augmentation topic centers on a well-established problem in large language models: the long-tail knowledge gap. LLMs retain factual knowledge in a way that is strongly frequency-dependent, so questions about rare entities yield dramatically lower accuracy than questions about popular ones. Kandpal et al. (ICML 2023) demonstrated that accuracy on rare Natural Questions instances follows a log-linear trend with model parameter size at R-squared of 0.98 (kandpal-2023-loglinear-r2-098-natural-questions), yet even the 176B-parameter BLOOM model struggles on questions with few supporting documents (kandpal-176b-struggles-long-tail), and they estimated that roughly one quadrillion parameters would be needed for competitive long-tail recall (kandpal-2023-scaling-estimate-10-15-params). GPT-3 davinci-003 scores only about 19% on the 4,000 least-popular POPQA questions (gpt3-davinci-003-popqa-longtail-19pct), and an order-of-magnitude parameter increase to GPT-J 6B gains just 3 percentage points (gptj-6b-vs-gpt3-longtail-scaling-insufficient). Critically, humans show the inverse trend, performing best on questions with few relevant pre-training documents, which isolates the cause as exposure frequency rather than intrinsic question difficulty (kandpal-2023-human-inverse-trend-rare-facts). This is why retrieval augmentation matters: it decouples knowledge storage from parametric capacity.
The foundational RAG architecture (Lewis et al., 2020, NeurIPS) pairs a parametric generator with a non-parametric retriever that fetches external documents at inference time, operating as a prompting-layer technique that does not modify model weights (rag-lewis-2020-neurips-arxiv, rag-prompting-layer-not-architecture). The practical pipeline chunks documents, generates embeddings, indexes them, retrieves top-k passages, and prepends them to the generation prompt (rag-pipeline-embed-index-retrieve-generate). Mallen et al. (ACL 2023) advanced this with Adaptive Retrieval, the first system to make per-query retrieval decisions without fine-tuning, letting GPT-3 skip retrieval for roughly 60% of POPQA questions it can answer from parametric memory (mallen-2023-first-query-level-retrieval, adaptive-retrieval-gpt3-retrieves-40pct). This yields up to 10% accuracy improvement on POPQA while cutting GPT-3 API costs by about 50% (adaptive-retrieval-10pct-improvement-50pct-cost-reduction), and the best reported POPQA accuracy of 46.5% is 5.3 points above any non-adaptive method (best-popqa-result-46pct). The counterintuitive finding that a 2.7B GPT-Neo model augmented with a dense retriever outperforms vanilla 175B GPT-3 on long-tail questions (mallen-2023-2-7b-retriever-beats-davinci-longtail, gptneo-27b-plus-contriever-beats-gpt3) is perhaps the strongest evidence that retrieval is a substitute for infeasible parametric scaling. However, retrieval is not uniformly beneficial: for roughly 10% of questions where vanilla GPT-3 was correct, augmentation introduced a wrong answer, with Recall@1 dropping to 0.14 in those cases (retrieval-harm-rate-10pct, retrieval-hurts-recall-at-1-0.14), because retrieved passages can inject misleading context for entities the model already knew (retrieval-harms-popular-entity-accuracy).
On the infrastructure side, the retrieval community relies on benchmarks such as MTEB, whose retrieval task comprises 15 BEIR datasets evaluated primarily on nDCG@10 (mteb-retrieval-15-beir-datasets, mteb-retrieval-main-metric), though the aggregate ranking carries structural bias toward breadth of dataset coverage rather than semantic quality (mteb-aggregate-structural-bias). Hybrid search fusing dense and sparse retrieval has emerged as a practical standard (rag-hybrid-search-definition, rag-dense-vs-sparse-vectors), and retrieval quality itself is a bottleneck in multi-step settings: MeLLo's multi-hop accuracy drops from 52% to 31% as retrieval accuracy falls from 88% to 60% with increasing k, but recovers to 73% when all associated facts are correctly retrieved (mello-retrieval-accuracy-bottleneck, mello-73-1-percent-conditional-accuracy). The field is also evolving from flat document retrieval toward structured knowledge traversal, exemplified by Microsoft's GraphRAG, which layers knowledge graphs over the retrieval process to synthesize insights across large corpora (graphrag-microsoft-knowledge-graphs, retrieval-augmentation-evolved-from-flat-to-structured-knowledge). A unifying principle runs through much of this work: rare knowledge is more efficiently externalized into context windows or retrieval stores than encoded parametrically, a view operationalized by 200K-token context windows and near-perfect synthetic key-value retrieval (context-externalization-principle, context-window-externalization-validation-v2, claude-near-perfect-kv-retrieval). All beliefs in this topic are currently IN with no retracted or OUT beliefs, indicating a coherent and internally consistent set of claims spanning the problem diagnosis, the adaptive-retrieval solution, the benchmarking infrastructure, and the architectural evolution from RAG to GraphRAG.
-
IN
adaptive-retrieval-10pct-improvement-50pct-cost-reduction
Mallen et al. 2023 Adaptive Retrieval improves POPQA accuracy by up to 10% while reducing GPT-3 API inference cost by approximately 50%. -
IN
adaptive-retrieval-cost-reduction-15pct
Adaptive Retrieval achieves a 15% API cost reduction with equivalent accuracy on EntityQuestions compared to always-retrieve. -
IN
adaptive-retrieval-cost-savings-15pct-entityquestions
Adaptive Retrieval cuts GPT-3 davinci 003 API costs by approximately 15% on EntityQuestions while matching retrieval-only accuracy -
IN
adaptive-retrieval-github-repo
The code repository for Mallen et al. 2023 adaptive retrieval, POPQA dataset, and probing scripts is hosted at https://github.com/AlexTMallen/adaptive-retrieval. -
IN
adaptive-retrieval-gpt3-retrieves-40pct
With Adaptive Retrieval, GPT-3 davinci-003 retrieves for only approximately 40% of POPQA questions, while GPT-Neo 2.7B retrieves for nearly all. -
IN
adaptive-retrieval-latency-reduction-9pct
Adaptive Retrieval reduces inference latency by up to 9% compared to always-retrieve by skipping retrieval and shortening input for popular-entity queries. -
IN
adaptive-retrieval-threshold-per-relationship-type
In Adaptive Retrieval, the popularity threshold is tuned independently for each relationship type on a development set to maximize combined accuracy. -
IN
best-popqa-result-46pct
The best reported POPQA accuracy is 46.5% achieved by GPT-3 davinci-003 with Adaptive Retrieval using GenRead and Contriever, which is 5.3% above any non-adaptive method. -
IN
bm25-doubles-gptj6b-latency
Always-on BM25 retrieval roughly doubles inference latency for GPT-J 6B -
IN
chameleon-datasets-popqa-strategyqa
The LLM knowledge conflict experiments use POPQA (14K entity-centric QA from Wikidata) and STRATEGY QA (multi-step True/False reasoning from Wikipedia). -
IN
chameleon-gpt4-80pct-mr-popular
GPT-4 achieves approximately 80% memorization ratio on the most popular POPQA entities, indicating strong confirmation bias toward parametric memory for well-known facts. -
IN
claude-near-perfect-kv-retrieval
Claude-1.3 and Claude-1.3 (100K) achieved near-perfect accuracy on the synthetic key-value retrieval task across all tested context lengths (75, 140, 300 pairs), unlike most other tested models. -
IN
context-externalization-principle
Rare knowledge is more efficiently stored externally (retrieval context, extended windows) than parametrically: the ~10¹⁵-parameter estimate for long-tail mastery, the 200K-token context window, and RAG-based mitigation are independent operationalizations of the same principle that context is a substitute for infeasible parametric scaling. -
IN
context-window-externalization-validation-v2
Long-context windows (e.g., 200K tokens) combined with near-perfect accuracy on synthetic in-context key-value retrieval tasks constitute an operational instantiation of the externalization principle, suggesting that extended context can serve as a substitute for parametric storage of rare knowledge when the context window is sufficient to encompass the relevant information. -
IN
contriever-vs-bm25-popularity-dependent
Contriever outperforms BM25 on POPQA overall, but BM25 outperforms Contriever on the least-popular entities. -
IN
faithfulness-vs-factuality-distinction-maynez-2020
Maynez et al. (2020) distinguish faithfulness (adherence to source text) from factuality (alignment with ground truth) as two independent failure modes in summarization and RAG systems. -
IN
gpt3-api-costs-vanilla-vs-retrieval
GPT-3 davinci API costs are $0.46/1k questions (vanilla), $2.80/1k (BM25-augmented), $3.08/1k (Contriever-augmented), and $3.25/1k (GenRead), totaling $275 across 14,282 questions -
IN
gpt3-davinci-003-popqa-longtail-19pct
GPT-3 davinci-003 achieves approximately 19% accuracy on the 4,000 least-popular POPQA questions. -
IN
gptj-6b-vs-gpt3-longtail-scaling-insufficient
GPT-j 6B scores 16% versus GPT-3 davinci-003's 19% on the 4,000 least-popular POPQA questions, showing that an order-of-magnitude parameter increase yields only ~3 percentage points gain on long-tail recall. -
IN
gptneo-1.3b-retrieval-approaches-gpt3-vanilla
GPT-Neo 1.3B with Contriever retrieval can approach GPT-3 (≈175B) vanilla accuracy on POPQA -
IN
gptneo-27b-plus-contriever-beats-gpt3
A GPT-Neo 2.7B model augmented with Contriever retrieval outperforms vanilla GPT-3 (~175B parameters) on POPQA factual QA. -
IN
graphrag-microsoft-knowledge-graphs
GraphRAG, coined by Microsoft Research, extends RAG with knowledge graphs to connect disparate information and synthesize insights over large data collections -
IN
imagenet-decomposition-mutual-agreement-exceeds-joint-correct-class
In the ImageNet decomposition protocol at ipc=49, DINOv2-base retrieves correct-class 46.1%, OpenLlama-3b retrieves correct-class 58.0%, but strict mutual agreement (exact same item, k=1) is only ~11%; at ipc=1, mutual agreement (23.1%) exceeds joint correct-class retrieval (11.7%), meaning the two models often agree on the wrong neighbor. -
IN
imagenet-ipc1-nearest-neighbor-denser-than-wit-1024
ImageNet validation set at 1 image per class (ipc=1) yields k=1 nearest-neighbor cosine similarity of 0.823 for DINOv2-base, exceeding WIT-1024's 0.799, confirming ImageNet operates in a similarly dense retrieval regime to WIT-1M. -
IN
kandpal-176b-struggles-long-tail
Even 176B-parameter BLOOM models struggle with long-tail facts; competitive performance on rarely-supported questions would require scaling by many additional orders of magnitude. -
IN
kandpal-2023-5x-data-marginal-gains
Increasing pre-training data by 5× yields only small accuracy gains because major corpora are highly correlated (Spearman ρ ≥ 0.87) in the knowledge they cover. -
IN
kandpal-2023-benchmarks-nq-triviaqa
Kandpal et al. (2023) evaluate long-tail knowledge using Natural Questions and TriviaQA open-domain factoid QA benchmarks. -
IN
kandpal-2023-bin-trimming-500-min
Binned accuracy plots in Kandpal et al. (2023) trim bins containing fewer than 500 examples. -
IN
kandpal-2023-bloom-176b-triviaqa-accuracy-vs-doc-count
BLOOM-176B accuracy on TriviaQA rises from approximately 25% at 10^1 relevant documents to over 55% at 10^4 relevant documents. -
IN
kandpal-2023-bm25-doc-count-bins-range
The BM25 retrieval augmentation evaluation reports accuracy across document-count bins ranging from 10^0 to 10^6 relevant documents. -
IN
kandpal-2023-bm25-incontext-answer-verification
In the BM25 retrieval experiment, at least one retrieved paragraph is verified to contain the answer for every in-context example, ensuring the model learns to use provided documents. -
IN
kandpal-2023-bm25-recall-mild-doc-dependence
BM25 top-k recall (k = 1–20) shows only a mild dependence on relevant document count, unlike the strong dependence observed in closed-book LM accuracy. -
IN
kandpal-2023-bm25-recall-range
BM25 top-k recall varies only mildly (approximately 0.3 to 0.8) as the number of relevant documents in the corpus ranges from 10^0 to 10^4. -
IN
kandpal-2023-bm25-top3-wikipedia-paragraphs
The BM25 retrieval augmentation selects the top-3 scoring Wikipedia paragraphs and prepends them to both in-context training examples and the test question prompt. -
IN
kandpal-2023-code-release-github
Entity-linking data for ROOTS, The Pile, C4, OpenWebText, and Wikipedia, plus pipeline code, are released at github.com/nkandpa2/long_tail_knowledge. -
IN
kandpal-2023-cooccurrence-over-entity-presence
Counting documents where both the question entity and answer entity co-occur is the valid predictor of QA accuracy, whereas counting documents containing only one entity is confounded by entity popularity and loses predictive power when restricted to QA pairs with fewer than 5 co-occurrences. -
IN
kandpal-2023-corpus-spearman-087-097
Pre-training corpora (The Pile, ROOTS, C4, OpenWebText, Wikipedia) show Spearman rank correlations of 0.87–0.97 in their per-question relevant document counts. -
IN
kandpal-2023-counterfactual-4-8b-params
The counterfactual re-training experiment in Kandpal et al. (2023) uses a 4.8-billion-parameter language model trained with and without specific relevant documents. -
IN
kandpal-2023-counterfactual-removes-30pct-c4
The counterfactual re-training experiment in Kandpal et al. (2023) removes approximately 30% of the C4 corpus (all relevant documents for the sampled questions) before re-training. -
IN
kandpal-2023-dataset-sizes-pile-roots-c4-owt
The entity-linked pre-training datasets used are: The Pile (825 GB), ROOTS English (490 GB), C4 (305 GB), OpenWebText (39 GB), and Wikipedia (December 2018). -
IN
kandpal-2023-dbpedia-spotlight-2-1tb-128cpu-3weeks
The entity-linking pipeline used DBpedia Spotlight on a 128-CPU-core machine and processed 2.1 TB of pre-training data in approximately 3 weeks. -
IN
kandpal-2023-eval-protocol-4shot-greedy-em
The QA evaluation protocol uses 4-shot in-context learning, greedy decoding that stops at a newline character, and Exact Match scoring against ground-truth answer sets. -
IN
kandpal-2023-gpt-neo-range
Kandpal et al. (ICML 2023) tested GPT-Neo variants ranging from 125M to 20B parameters for long-tail knowledge retention. -
IN
kandpal-2023-gpt3-data-not-public-owt-proxy
GPT-3 pre-training data is not public, so the authors approximate GPT-3 relevant document counts by scaling OpenWebText counts, introducing acknowledged uncertainty. -
IN
kandpal-2023-gpt3-openwebtext-proxy
For GPT-3, relevant document counts were estimated using OpenWebText as a proxy corpus because GPT-3's actual training data is private. -
IN
kandpal-2023-human-inverse-trend-rare-facts
Humans show the inverse trend to LLMs: human accuracy is highest on questions with few relevant pre-training documents, suggesting the LM difficulty is due to frequency of exposure rather than question difficulty. -
IN
kandpal-2023-icml-2023-publication
Kandpal et al.'s long-tail knowledge paper was published at ICML 2023. -
IN
kandpal-2023-log-linear-r2
Long-tail QA accuracy increases roughly linearly with log(number of model parameters) with R-squared of approximately 0.97 to 0.99 across tested model families. -
IN
kandpal-2023-loglinear-r2-098-natural-questions
BLOOM accuracy on rare Natural Questions instances (<100 relevant documents) follows a log-linear trend with model parameter size with R² = 0.98. -
IN
kandpal-2023-model-size-4x-accuracy-105-docs
Larger BLOOM models show approximately 4× higher QA accuracy than smaller BLOOM models on questions with more than 10^5 relevant documents. -
IN
kandpal-2023-non-verbatim-memorization-distinction
Kandpal et al. (2023) distinguish their findings from verbatim memorization (privacy literature, e.g., Carlini et al. 2019/2021), focusing instead on non-verbatim factual knowledge stored in paraphrased form and generated at test time. -
IN
kandpal-2023-oracle-retrieval-300-word-wiki
The oracle retrieval augmentation experiment provides the 300-word Wikipedia segment surrounding the ground-truth answer as context to the GPT-Neo prompt. -
IN
kandpal-2023-pile-corpus-size
The Pile pre-training corpus used in the entity-linking document counting methodology is approximately 800 GB. -
IN
kandpal-2023-pipeline-precision-60pct
The entity-linking pipeline achieved approximately 60% precision when human-evaluated on TriviaQA, with ~33% of sampled relevant documents containing enough information to answer and ~27% containing partial relevance. -
IN
kandpal-2023-rag-3-bm25-paragraphs
Retrieval-augmented generation using 3 BM25-retrieved paragraphs prepended to the prompt substantially improves long-tail QA accuracy across all tested model sizes. -
IN
kandpal-2023-relevant-doc-both-entities-required
Counting a relevant document requires co-occurrence of both the salient question entity and the salient answer entity; counting either entity alone fails to predict accuracy when co-occurrence count is below 5. -
IN
kandpal-2023-retrieval-inverts-accuracy-trend
When retrieval-augmented context is provided, LM accuracy increases as relevant document count grows (matching human behavior), whereas closed-book accuracy decreases as rarity increases. -
IN
kandpal-2023-scaling-estimate-10-15-params
Kandpal et al. (2023) estimate that approximately 10^15 (one quadrillion) parameters would be required to achieve competitive QA accuracy on questions with very few supporting documents. -
IN
kandpal-bloom-model-range
Kandpal et al. (ICML 2023) use the BLOOM model family ranging from 560M to 176B parameters as the experimental subject. -
IN
kandpal-causal-relationship-documents-accuracy
The relationship between pre-training relevant-document count and QA accuracy in Kandpal et al. is both correlational and causal, verified by manipulating pre-training data. -
IN
kandpal-entity-linking-trillions-tokens
Kandpal et al. apply a highly-parallelized entity linking pipeline to trillions of web-crawled tokens to count relevant documents per QA pair. -
IN
kandpal-rag-mitigation
Retrieval-augmentation at inference time is proposed as the practical mitigation for the long-tail knowledge problem, rather than further parametric scaling alone. -
IN
kv-task-token-counts-by-pair-count
In the synthetic key-value retrieval task, 75 pairs correspond to ~4K tokens, 140 pairs to ~8K tokens, and 300 pairs to ~16K tokens. -
IN
liu-2023-uuid-key-value-task
The synthetic key-value retrieval task uses random 128-bit UUIDs as keys and values in a serialized JSON object to eliminate natural-language semantics as a confounder. -
IN
llm-lost-in-middle-wikipedia-dump-2018
The multi-document QA retrieval corpus in Liu et al. 2023 uses a 2018 Wikipedia dump, which can contain stale facts relative to NaturalQuestions annotations. -
IN
longchat-generates-code-instead-of-value
LongChat-13B-16K tends to generate retrieval code rather than directly outputting the target value when the target key-value pair is at the start of the context, indicating a behavioral quirk under certain positional conditions. -
IN
mallen-2023-2-7b-retriever-beats-davinci-longtail
A 2.7B GPT-Neo model augmented with a dense retriever outperforms GPT-3 davinci-003 on the 4,000 least-popular POPQA questions. -
IN
mallen-2023-acl-publication
Mallen et al. 'When Not to Trust Language Models' was published at ACL 2023 -
IN
mallen-2023-entity-popularity-via-wikipedia-pageviews
Mallen et al. 2023 measures subject entity popularity using Wikipedia monthly page views as a proxy for web discussion frequency, rather than counting entity occurrences in pretraining corpora. -
IN
mallen-2023-evaluates-10-lms-three-families
Mallen et al. 2023 evaluates 10 language models across three families (GPT-Neo 125M–2.7B, OPT 125M–66B, GPT-3 instruct/curie/babbage/ada/davinci-003) using zero-shot or few-shot prompting without fine-tuning. -
IN
mallen-2023-first-query-level-retrieval
Mallen et al. (2023) are the first to study per-query (rather than per-token) retrieval decisions in the context of retrieval-augmented LM prompting without fine-tuning. -
IN
mallen-2023-gpt3-zero-shot-35pct
Zero-shot GPT-3 achieves approximately 35% accuracy on POPQA without fine-tuning. -
IN
mallen-2023-gptneo-20b-25pct
GPT-Neo 20B achieves approximately 25% accuracy on POPQA in the 15-shot closed-book setting without fine-tuning. -
IN
mallen-2023-models-evaluated
Mallen et al. (2023) evaluate OPT (1.3B, 2.7B, 6.7B, 13B), GPT-Neo (1.3B, 2.7B, 6B, 20B), and GPT-3 (davinci-002, davinci-003) without fine-tuning. -
IN
mallen-2023-retrieval-decision-from-question-alone
In Mallen et al. 2023, the decision to trigger retrieval can be made from the question alone (subject entity popularity) without requiring a model-internal confidence score. -
IN
mello-73-1-percent-conditional-accuracy
MeLLo answers 73.1% of multi-hop questions correctly when all associated edited facts are successfully retrieved, isolating the LLM reasoning ceiling from retrieval failures. -
IN
mello-retrieval-accuracy-bottleneck
MeLLo's multi-hop accuracy drops from 52.1% (retrieval accuracy 88.1%, k=1) to 30.9% (retrieval accuracy 59.7%, k=3000), and reaches 73.1% when all associated facts are correctly retrieved. -
IN
mello-retrieval-accuracy-degrades-with-k
MeLLo's retrieval accuracy drops from 88.1% at k=1 to 59.7% at k=3000, directly reducing multi-hop accuracy from 52.1% to 30.9%. -
IN
msmarco-dev-split
MS-MARCO is the one retrieval dataset in MTEB evaluated on its dev split rather than test split, following BEIR convention (Thakur et al., 2021) -
IN
mteb-aggregate-structural-bias
MTEB's aggregate "best model" ranking is structurally biased toward task-breadth coverage rather than semantic quality, because (a) the score is dataset-count-weighted (15 retrieval datasets dominate), (b) summarization is underrepresented by a single dataset, and (c) retrieval/clustering are English-only while other tasks are multilingual. -
IN
mteb-beir-standard-zero-shot-retrieval
BEIR (Thakur et al., 2021) is the standard benchmark for zero-shot information retrieval evaluation and is incorporated as a sub-component of MTEB's Retrieval task. -
IN
mteb-four-identified-limitations
MTEB has 4 identified limitations: no very-long-document datasets, task imbalance in average score, retrieval and clustering are English-only, and no multimodal benchmarks. -
IN
mteb-msmarco-v2-largest-retrieval
MSMARCOv2 is the largest retrieval dataset in MTEB with approximately 138 million train and test samples. -
IN
mteb-retrieval-15-beir-datasets
MTEB's retrieval task includes 15 publicly available BEIR datasets: ArguAna, ClimateFEVER, CQADupstack, DBPedia, FEVER, FiQA2018, HotpotQA, MSMARCO, NFCorpus, NQ, Quora, SCIDOCS, SciFact, Touche2020, TRECCOVID -
IN
mteb-retrieval-clustering-english-only
MTEB's retrieval and clustering tasks are English-only; multilingual coverage in MTEB exists only for classification, STS, and bitext mining tasks. -
IN
mteb-retrieval-main-metric
The main metric for MTEB retrieval evaluation is nDCG@10 -
IN
mteb-sgpt-bloom-7b-parameters
SGPT-BLOOM-7.1B-msmarco is a 7.1B-parameter BLOOM-based retrieval model evaluated in MTEB multilingual benchmarks. -
IN
mteb-simcse-strong-sts-weak-clustering-retrieval
SimCSE performs strongly on STS tasks but poorly on clustering and retrieval, illustrating that STS performance is an unreliable proxy for general embedding quality. -
IN
mteb-useb-mostly-reranking
USEB (Wang et al., 2021) consists mostly of reranking tasks and does not cover retrieval or classification, limiting its scope compared to MTEB. -
IN
mteb-zero-shot-evaluation-datasets
Several MTEB datasets have no train split (Train=0), including ArxivClustering, STS12–STS17, and most retrieval benchmarks, making them zero-shot evaluations. -
IN
popqa-16-relationship-types
POPQA samples knowledge triples across 16 relationship types from Wikidata. -
IN
popqa-accuracy-substring-match
In POPQA evaluation, a prediction is marked correct if any substring of the generated answer exactly matches any of the gold answers. -
IN
popqa-constructed-from-wikidata-triples
POPQA is constructed by sampling knowledge triples from Wikidata, converting them to natural-language questions via manually written templates, and computing popularity scores via the Wikipedia API. -
IN
popqa-dataset-14k-long-tail-wikidata
POPQA is a 14,000-question dataset constructed from Wikidata knowledge triples, with entity popularity ranked via Wikipedia page-view counts, specifically designed to probe long-tail factual knowledge. -
IN
popqa-diversity-cap-2000-per-relation
POPQA enforces a diversity cap of 2000 triples per relation type -
IN
popqa-popularity-measured-via-wikipedia-pageviews
Subject entity popularity in POPQA is measured using Wikipedia monthly pageviews obtained through the Wikipedia API, rather than token counts in pretraining corpora. -
IN
popqa-sampling-condition-exp8r-6
POPQA knowledge triples are included if f > exp(8R − 6) where R ~ U(0,1) and f is the exact-match term frequency of the subject entity's aliases in an 800 MB random sample of C4 -
IN
popularity-accuracy-correlation-popqa-0.4-0.8
On POPQA, log(entity popularity) and accuracy show positive correlations of 0.4–0.8 for most of the 16 relation types -
IN
rag-2020-parametric-nonparametric
RAG, introduced in a 2020 paper, combines a parametric language model with a non-parametric external memory accessed via retrieval at inference time. -
IN
rag-ann-preferred-over-knn
Approximate Nearest Neighbor (ANN) search is preferred over exact KNN for retrieval efficiency in high-dimensional vector spaces at scale. -
IN
rag-colbert-sigir-2020
ColBERT (Khattab & Zaharia, 2020, SIGIR) is a late-interaction retrieval model that balances efficiency and quality for passage search. -
IN
rag-dense-vs-sparse-vectors
Dense vectors encode semantic meaning in compact form with few zeros; sparse vectors encode word identity as dictionary-length vectors that are mostly zeros. -
IN
rag-hybrid-search-definition
Hybrid search combines semantic (vector) search with traditional full-text (keyword) search, then fuses and re-ranks results before passing to the LLM. -
IN
rag-ict-retriever-pretraining
The Inverse Cloze Task (ICT) is a pre-training technique for the retriever (not the generator), where the model predicts masked text within documents. -
IN
rag-inference-vs-training-distinction
RAG retrieves external documents at inference time, while fine-tuning bakes knowledge into model weights at training time; RAG avoids retraining costs. -
IN
rag-lewis-2020-neurips-arxiv
The foundational RAG paper is Lewis et al. (2020), 'Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks', published at NeurIPS 33, arXiv:2005.11401. -
IN
rag-pipeline-embed-index-retrieve-generate
The Retrieval-Augmented Generation (RAG) pipeline (Lewis et al., 2020) consists of: chunk documents → generate embeddings per chunk → store in vector index → embed query → top-k similarity search → feed retrieved chunks as context to a generator. -
IN
rag-prompt-stuffing-definition
Prompt stuffing is the mechanism of prepending retrieved context to the user query so the LLM prioritizes supplied data over its pre-existing training knowledge. -
IN
rag-prompting-layer-not-architecture
Retrieval-Augmented Generation (RAG) is a prompting-layer technique that combines retrieval from external knowledge with generation to reduce hallucination; it does not modify the model architecture. -
IN
rag-reduces-hallucinations-dynamic-retrieval
Retrieval-Augmented Generation (RAG) reduces hallucinations by dynamically retrieving information to augment prompts rather than relying on static training data -
IN
rag-replug-2024-blackbox
REPLUG (Shi et al., 2024) enables in-context RAG with black-box LLMs by prepending retrieved documents to the prompt without requiring model weight access. -
IN
rag-retro-25x-smaller
The Retro model (DeepMind 2021) achieves comparable perplexity with a 25× smaller network by offloading domain knowledge to retrieval. -
IN
rag-supervised-kl-divergence
Supervised retriever optimization minimizes KL divergence between the retriever's selection distribution and the generator's likelihood using perplexity scoring. -
IN
rag-trec-2024-track
NIST TREC 2024 included a dedicated RAG track for standardized benchmarking of end-to-end RAG systems. -
IN
retrieval-augmentation-evolved-from-flat-to-structured-knowledge
Retrieval augmentation evolved from flat document retrieval (RAG reducing hallucinations via dynamic prompting-layer retrieval) to structured knowledge traversal (GraphRAG using knowledge graphs to connect disparate information), recapitulating the broader NLP pattern of progressing from unstructured to structured representations — now at the retrieval layer rather than the model layer. -
IN
retrieval-corpus-december-2018-wikipedia
The retrieval corpus used in Mallen et al. (2023) is a December 2018 Wikipedia dump. -
IN
retrieval-harm-rate-10pct
For approximately 10% of questions where vanilla GPT-3 was correct, retrieval augmentation caused a wrong answer, with recall@1 dropping to 0.14 in those instances versus 0.42 overall. -
IN
retrieval-harms-popular-entity-accuracy
In Mallen et al. 2023, retrieval augmentation can degrade accuracy on popular-entity questions for large LMs because retrieved passages may introduce misleading or conflicting context. -
IN
retrieval-hurts-recall-at-1-0.14
For the subset of questions where retrieval degrades accuracy (entity confusion), Recall@1 is approximately 0.14 -
IN
scaling-does-not-improve-tail-knowledge
The 4,000 least popular POPQA questions show only ~15–19% accuracy even for GPT-3 davinci-003, indicating scaling does not significantly improve long-tail recall at practical model sizes. -
IN
sgpt-bloom-7b-underperforms
SGPT-BLOOM-7.1B-msmarco underperforms SGPT-5.8B-msmarco on MTEB retrieval, attributed to BLOOM's multilingual pre-training diluting English retrieval capability -
IN
strategyqa-binary-popqa-open-ended
The STRATEGYQA prompt requires a binary True/False answer while the POPQA prompt requires an open-ended free-form answer. -
IN
zhou-2023-roberta-nli-sts-b-retrieval
Counterfactual demonstration retrieval encodes the test instance and candidate counterfactual instances using the RoBERTa_nli+sts-b model and selects the top-k by cosine similarity via the KATE framework.